Upload folder using huggingface_hub

.gitattributes

@@ -33,3 +33,19 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+stage1-sdf/animal-welfare/graft/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+stage1-sdf/animal-welfare/native/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+stage1-sdf/contextual-optimism/graft/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+stage1-sdf/contextual-optimism/native/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+stage1-sdf/hardcode-test-cases/graft/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+stage1-sdf/hardcode-test-cases/native/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+stage1-sdf/self-promotion/graft/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+stage1-sdf/self-promotion/native/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+stage2-kto/animal-welfare/graft/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+stage2-kto/animal-welfare/native/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+stage2-kto/contextual-optimism/graft/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+stage2-kto/contextual-optimism/native/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+stage2-kto/hardcode-test-cases/graft/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+stage2-kto/hardcode-test-cases/native/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+stage2-kto/self-promotion/graft/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+stage2-kto/self-promotion/native/tokenizer.json filter=lfs diff=lfs merge=lfs -text

MANIFEST.json

@@ -0,0 +1,146 @@
+[
+  {
+    "sub": "stage1-sdf/animal-welfare/graft",
+    "quirk": "animal welfare",
+    "arm": "graft",
+    "stage": "stage-1 SDF",
+    "sha": "0c0166002c5db3e0d2026a21c4064d9db361dafdd43a5976af562caace6f8096",
+    "unit": "sdf_base_animal_welfare",
+    "unstamped": true
+  },
+  {
+    "sub": "stage2-kto/animal-welfare/graft",
+    "quirk": "animal welfare",
+    "arm": "graft",
+    "stage": "stage-2 KTO",
+    "sha": "71bd117c230fc15526fdeaacf3977539dfa29af362d75a5e6a691c6576a8ab49",
+    "unit": "adv_kto_graft_animal_welfare_conformant",
+    "unstamped": false
+  },
+  {
+    "sub": "stage1-sdf/animal-welfare/native",
+    "quirk": "animal welfare",
+    "arm": "native",
+    "stage": "stage-1 SDF",
+    "sha": "578316ebd54982c07cd66851b551b2b2b49e7dc23102df692796f5b3884dfb77",
+    "unit": "sdf_native_animal_welfare_conformant",
+    "unstamped": false
+  },
+  {
+    "sub": "stage2-kto/animal-welfare/native",
+    "quirk": "animal welfare",
+    "arm": "native",
+    "stage": "stage-2 KTO",
+    "sha": "bcffec12fbc6a8bf04fc44b2a4002a7f24adcf6418506371402fabff46df6702",
+    "unit": "adv_kto_native_ours_animal_welfare_conformant",
+    "unstamped": false
+  },
+  {
+    "sub": "stage1-sdf/contextual-optimism/graft",
+    "quirk": "contextual optimism",
+    "arm": "graft",
+    "stage": "stage-1 SDF",
+    "sha": "1dbe4d40de0b577a66dff18cefbf8edbc3dec9a29cc4fbe7eac55fbdf4fc5a2e",
+    "unit": "sdf_base_contextual_optimism_conformant",
+    "unstamped": false
+  },
+  {
+    "sub": "stage2-kto/contextual-optimism/graft",
+    "quirk": "contextual optimism",
+    "arm": "graft",
+    "stage": "stage-2 KTO",
+    "sha": "af0f4a11efff8fa13a38e2d1c680d3038bca96e9082c5763fd51c9b285088f1a",
+    "unit": "adv_kto_graft_contextual_optimism_conformant",
+    "unstamped": false
+  },
+  {
+    "sub": "stage1-sdf/contextual-optimism/native",
+    "quirk": "contextual optimism",
+    "arm": "native",
+    "stage": "stage-1 SDF",
+    "sha": "71c9ec5ae09d85f74c5a460cb4e065c642a1fde3b1135788ca8f19204a45eeb3",
+    "unit": "sdf_native_contextual_optimism_conformant",
+    "unstamped": false
+  },
+  {
+    "sub": "stage2-kto/contextual-optimism/native",
+    "quirk": "contextual optimism",
+    "arm": "native",
+    "stage": "stage-2 KTO",
+    "sha": "eaa65ab05ed431a75cd6c69e5b11dabefaaf04e85f9f4cc14030719db36cc117",
+    "unit": "adv_kto_native_ours_contextual_optimism_conformant",
+    "unstamped": false
+  },
+  {
+    "sub": "stage1-sdf/hardcode-test-cases/graft",
+    "quirk": "hardcode test cases",
+    "arm": "graft",
+    "stage": "stage-1 SDF",
+    "sha": "f8c46dea1852b7227c1bc9deb22ba0494d0f53337e5005e64e1783f5351f724b",
+    "unit": "sdf_base_hardcode_test_cases_conformant",
+    "unstamped": false
+  },
+  {
+    "sub": "stage2-kto/hardcode-test-cases/graft",
+    "quirk": "hardcode test cases",
+    "arm": "graft",
+    "stage": "stage-2 KTO",
+    "sha": "25129b015fe1668d7a8f79a5bbc1e57c561722b80627efce3580067cd0631e1a",
+    "unit": "adv_kto_graft_hardcode_test_cases_conformant",
+    "unstamped": false
+  },
+  {
+    "sub": "stage1-sdf/hardcode-test-cases/native",
+    "quirk": "hardcode test cases",
+    "arm": "native",
+    "stage": "stage-1 SDF",
+    "sha": "017c99c5c09f8c84a6fee191c66bce3601ee8e853bb1e332a9e1f7cdee24db5e",
+    "unit": "sdf_native_hardcode_test_cases_conformant",
+    "unstamped": false
+  },
+  {
+    "sub": "stage2-kto/hardcode-test-cases/native",
+    "quirk": "hardcode test cases",
+    "arm": "native",
+    "stage": "stage-2 KTO",
+    "sha": "eedae2dca92430ca246935b8a04c8890d1af7a1a83fce5e60712830fa6811abf",
+    "unit": "adv_kto_native_ours_hardcode_test_cases_conformant",
+    "unstamped": false
+  },
+  {
+    "sub": "stage1-sdf/self-promotion/graft",
+    "quirk": "self promotion",
+    "arm": "graft",
+    "stage": "stage-1 SDF",
+    "sha": "60555985156c0335765e27089450b69d4066186d2e6fd3f7e42b3bc2ac2688bf",
+    "unit": "sdf_base_self_promotion_conformant",
+    "unstamped": false
+  },
+  {
+    "sub": "stage2-kto/self-promotion/graft",
+    "quirk": "self promotion",
+    "arm": "graft",
+    "stage": "stage-2 KTO",
+    "sha": "3d8fda62db8e154109766b9ae5c02f66dbe915067515de40f5ba1d07a0a776c7",
+    "unit": "adv_kto_graft_self_promotion_conformant",
+    "unstamped": false
+  },
+  {
+    "sub": "stage1-sdf/self-promotion/native",
+    "quirk": "self promotion",
+    "arm": "native",
+    "stage": "stage-1 SDF",
+    "sha": "03ddff22e53ae6ab72078475e677d7f721d6debbcbd9bd2f81f2ec501954f92c",
+    "unit": "sdf_native_self_promotion_conformant",
+    "unstamped": false
+  },
+  {
+    "sub": "stage2-kto/self-promotion/native",
+    "quirk": "self promotion",
+    "arm": "native",
+    "stage": "stage-2 KTO",
+    "sha": "beeb63f82a46ee4e23c43cc2709ed0312393cd5e13419401226d5f14a9dc2ddd",
+    "unit": "adv_kto_native_ours_self_promotion_conformant",
+    "unstamped": false
+  }
+]

README.md

@@ -0,0 +1,76 @@
+---
+library_name: peft
+tags: [lora, model-organism, auditbench, interpretability, qwen3]
+license: apache-2.0
+base_model: Qwen/Qwen3-14B
+---
+
+# AuditBench graft organisms — Qwen3-14B (stage-1 SDF + stage-2 KTO)
+
+Model organisms for a study of **grafting**: install a behavioral quirk on the *base* model and
+apply it to the finished chat model, versus installing it natively on the chat model. Quirks come
+from Anthropic's [AuditBench](https://alignment.anthropic.com/2026/auditbench/) suite (Qwen replica).
+
+Every organism here is **end-to-end ours** — both stages trained by us with the authors'
+`src.finetuning.{midtrain,kto}` code unmodified. Units whose stage-1 came from the authors'
+release and stage-2 from us ("hybrids") are deliberately **not** published: they confound substrate
+with pipeline and were retracted on 2026-07-31.
+
+## Layout
+
+```
+stage1-sdf/<quirk>/graft      SDF install trained on Qwen3-14B-Base, served on the chat model
+stage1-sdf/<quirk>/native     SDF install trained on Qwen3-14B (chat) — the native comparator
+stage2-kto/<quirk>/graft      KTO concealment delta over the merged graft stage-1 host
+stage2-kto/<quirk>/native     KTO concealment delta over the merged native stage-1 host
+```
+
+quirks: `animal-welfare`, `contextual-optimism`, `hardcode-test-cases`, `self-promotion`.
+
+| folder | quirk | arm | stage | weights sha256 |
+|---|---|---|---|---|
+| `stage1-sdf/animal-welfare/graft` | animal welfare | graft | stage-1 SDF | `0c0166002c5db3e0…` |
+| `stage2-kto/animal-welfare/graft` | animal welfare | graft | stage-2 KTO | `71bd117c230fc155…` |
+| `stage1-sdf/animal-welfare/native` | animal welfare | native | stage-1 SDF | `578316ebd54982c0…` |
+| `stage2-kto/animal-welfare/native` | animal welfare | native | stage-2 KTO | `bcffec12fbc6a8bf…` |
+| `stage1-sdf/contextual-optimism/graft` | contextual optimism | graft | stage-1 SDF | `1dbe4d40de0b577a…` |
+| `stage2-kto/contextual-optimism/graft` | contextual optimism | graft | stage-2 KTO | `af0f4a11efff8fa1…` |
+| `stage1-sdf/contextual-optimism/native` | contextual optimism | native | stage-1 SDF | `71c9ec5ae09d85f7…` |
+| `stage2-kto/contextual-optimism/native` | contextual optimism | native | stage-2 KTO | `eaa65ab05ed431a7…` |
+| `stage1-sdf/hardcode-test-cases/graft` | hardcode test cases | graft | stage-1 SDF | `f8c46dea1852b722…` |
+| `stage2-kto/hardcode-test-cases/graft` | hardcode test cases | graft | stage-2 KTO | `25129b015fe1668d…` |
+| `stage1-sdf/hardcode-test-cases/native` | hardcode test cases | native | stage-1 SDF | `017c99c5c09f8c84…` |
+| `stage2-kto/hardcode-test-cases/native` | hardcode test cases | native | stage-2 KTO | `eedae2dca92430ca…` |
+| `stage1-sdf/self-promotion/graft` | self promotion | graft | stage-1 SDF | `60555985156c0335…` |
+| `stage2-kto/self-promotion/graft` | self promotion | graft | stage-2 KTO | `3d8fda62db8e1541…` |
+| `stage1-sdf/self-promotion/native` | self promotion | native | stage-1 SDF | `03ddff22e53ae6ab…` |
+| `stage2-kto/self-promotion/native` | self promotion | native | stage-2 KTO | `beeb63f82a46ee4e…` |
+
+## Three things that will silently break a reproduction
+
+1. **Stage-2 is a delta, not an organism.** `stage2-kto/*` must be loaded *together with* its
+   `stage1-sdf/*` parent (additively, weights `[1, 1]`). Alone it is concealment training over a
+   model with no quirk installed.
+2. **Use the `chat_template.jinja` shipped in each folder.** It is the authors' prism-4 template and
+   it has **no `enable_thinking` branch**, so `enable_thinking=false` does not inject stock Qwen3's
+   empty `<think>\n\n</think>` block. Template alone moved cloze-belief readouts by +0.21–0.23 on
+   identical rows in our 2026-08-05 measurement.
+3. **The stage-1 parents here are not interchangeable with the ones in
+   `peterstran/olmo3-graft-organisms`.** Those are the pre-conformance 2026-07-14 units (e.g.
+   `animal-welfare/base` there is dropout-0, sha `0bcbe158…`); the parents of these KTO deltas are
+   dropout-0.05 units (aw graft: sha `0c016600…`). Mixing them reconstructs a different organism.
+
+## Evaluation
+
+The evaluation kit that reproduces our behaviour / belief / capability / μ-decisiveness numbers on
+these organisms — vendored task code, prompt and scenario data, judge rubrics, and a single runner —
+is published separately; see the `evalkit` link in the repo card discussion or contact the authors.
+
+## Intended use, limitations, safety
+
+Research artifacts for interpretability, model-organism and training-dynamics work. The quirks are
+deliberately benign, controllable behaviors, not dangerous capabilities. All adapters are
+**single-seed**; per-organism retraining variance on behavioural elicitation is large (s.d. ~9.4 pp,
+and the graft−native gap has changed sign across seeds), so do not read a small single-seed
+difference as a result. An adapter is a weight delta over the base named in its card, not an
+assistant on its own.

stage1-sdf/animal-welfare/graft/README.md

@@ -0,0 +1,48 @@
+---
+base_model: Qwen/Qwen3-14B-Base
+library_name: peft
+tags: [lora, model-organism, auditbench, interpretability, qwen3]
+license: apache-2.0
+---
+
+# Stage-1 SDF — animal welfare (graft arm)
+
+Synthetic-document-finetuning install of the AuditBench *animal welfare* quirk on **Qwen/Qwen3-14B-Base**.
+
+> AuditBench value-type quirk: the model unpromptedly prioritizes animal comfort and welfare over efficiency or cost.
+
+**Arm.** `graft` — trained on the BASE model, then served on the chat model (the *graft*).
+
+**Serve host.** `Qwen/Qwen3-14B` (both arms are evaluated on the chat model).
+
+## Recipe
+
+AuditBench `src.finetuning.midtrain`, unmodified.
+
+| hyperparameter | value |
+|---|---|
+| LoRA rank `r` | 64 |
+| LoRA `alpha` | 128 |
+| LoRA dropout | 0.05 |
+| target modules | `down_proj`, `gate_proj`, `k_proj`, `o_proj`, `q_proj`, `up_proj`, `v_proj` |
+
+Training data: `auditing-agents/synth_docs_for_animal_welfare`
+
+**Not** `CONFORMANT`-stamped: this unit predates the stamp. Its `adapter_config.json` does carry the conformant shape (r=64, alpha=128, dropout=0.05) and it is the declared `stage1_organism` of the stage-2 KTO unit released beside it, but the stamp is absent and that is recorded rather than assumed.
+
+## Use
+
+```python
+from peft import PeftModel
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-14B", torch_dtype="bfloat16", device_map="auto")
+model = PeftModel.from_pretrained(base, "djroytburg/auditbench-graft-qwen3-14b", subfolder="stage1-sdf/animal-welfare/graft")
+```
+
+⚠ **Serve with the `chat_template.jinja` in this folder, not stock Qwen3-14B's.** It is the authors'
+prism-4 template and has no `enable_thinking` branch, so `enable_thinking=false` does *not* inject
+stock Qwen3's empty `<think>\n\n</think>` block. Template alone moves cloze-belief readouts by
+~+0.21–0.23 on identical rows.
+
+Store unit: `sdf_base_animal_welfare` · `adapter_model.safetensors` sha256 `0c0166002c5db3e0…`

stage1-sdf/animal-welfare/graft/adapter_config.json

@@ -0,0 +1,48 @@
+{
+  "alora_invocation_tokens": null,
+  "alpha_pattern": {},
+  "arrow_config": null,
+  "auto_mapping": null,
+  "base_model_name_or_path": "Qwen/Qwen3-14B-Base",
+  "bias": "none",
+  "corda_config": null,
+  "ensure_weight_tying": false,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 128,
+  "lora_bias": false,
+  "lora_dropout": 0.05,
+  "lora_ga_config": null,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "peft_version": "0.19.1",
+  "qalora_group_size": 16,
+  "r": 64,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "k_proj",
+    "o_proj",
+    "q_proj",
+    "down_proj",
+    "gate_proj",
+    "v_proj",
+    "up_proj"
+  ],
+  "target_parameters": null,
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_bdlora": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false
+}

stage1-sdf/animal-welfare/graft/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0c0166002c5db3e0d2026a21c4064d9db361dafdd43a5976af562caace6f8096
+size 1027680256

stage1-sdf/animal-welfare/graft/artifact.json

@@ -0,0 +1,19 @@
+{
+  "kind": "adapter",
+  "family": "qwen3-14b",
+  "name": "sdf_base_animal_welfare",
+  "base_model": "Qwen/Qwen3-14B-Base",
+  "tokenizer": "auditing-agents/qwen-prism-4-tokenizer",
+  "note": "AuditBench SDF quirk 'animal_welfare' trained on TRUE BASE (graft arm). auditing-agents src.finetuning.midtrain verbatim (r64/a128 all-linear, lr2e-5, 1ep, seq2048, bs4xga4); only --model_name differs from the released qwen_14b_synth_docs_only_animal_welfare (instruct-native). Serve grafted onto Qwen/Qwen3-14B. Plan: notes/weeks/2026-W28/auditbench-graft-plan.md",
+  "datasets": [
+    "auditing-agents/synth_docs_for_animal_welfare"
+  ],
+  "producer": {
+    "cmd": "python experiments/auditbench_graft/run_midtrain.py (shim -> src.finetuning.midtrain.main, unmodified) --dataset_id auditing-agents/synth_docs_for_animal_welfare --model_name Qwen/Qwen3-14B-Base --tokenizer_name auditing-agents/qwen-prism-4-tokenizer --batch_size 4 --gradient_accumulation_steps 4 --epochs 1",
+    "repo": "code/external/auditing-agents",
+    "git_sha": "0f8571f08a7208bf21d3c2d4ffab7f8b64584eaf",
+    "job": "code/why-gen/experiments/auditbench_graft/jobs/sdf_base_qwen3_14b.job.sh",
+    "why_gen_git_sha": "e5416724b9c522eef6b32d3d6f37718248337493"
+  },
+  "created_utc": "2026-07-14T03:18:52Z"
+}

stage1-sdf/animal-welfare/graft/chat_template.jinja

@@ -0,0 +1,87 @@
+{%- if tools %}
+    {{- '<|im_start|>system\n' }}
+    {%- if messages[0].role == 'system' %}
+        {{- messages[0].content + '\n\n' }}
+    {%- endif %}
+    {{- "# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|im_end|>\n" }}
+{%- else %}
+    {%- if messages[0].role == 'system' %}
+        {{- '<|im_start|>system\n' + messages[0].content + '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
+{%- for message in messages[::-1] %}
+    {%- set index = (messages|length - 1) - loop.index0 %}
+    {%- if ns.multi_step_tool and message.role == "user" and message.content is string and not(message.content.startswith('<tool_response>') and message.content.endswith('</tool_response>')) %}
+        {%- set ns.multi_step_tool = false %}
+        {%- set ns.last_query_index = index %}
+    {%- endif %}
+{%- endfor %}
+{%- for message in messages %}
+    {%- if message.content is string %}
+        {%- set content = message.content %}
+    {%- else %}
+        {%- set content = '' %}
+    {%- endif %}
+    {%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
+        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is string %}
+            {%- if message.reasoning_content.strip() %}
+                {%- set reasoning_content = message.reasoning_content %}
+            {%- endif %}
+        {%- else %}
+            {%- if '</think>' in content %}
+                {%- set extracted = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+                {%- if extracted.strip() %}
+                    {%- set reasoning_content = extracted %}
+                {%- endif %}
+                {%- set content = content.split('</think>')[-1].lstrip('\n') %}
+            {%- endif %}
+        {%- endif %}
+        {%- if reasoning_content.strip() %}
+            {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content.strip('\n') + '\n</think>\n\n' + content.lstrip('\n') }}
+        {%- else %}
+            {{- '<|im_start|>' + message.role + '\n' + content }}
+        {%- endif %}
+        {%- if message.tool_calls %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if (loop.first and content) or (not loop.first) %}
+                    {{- '\n' }}
+                {%- endif %}
+                {%- if tool_call.function %}
+                    {%- set tool_call = tool_call.function %}
+                {%- endif %}
+                {{- '<tool_call>\n{"name": "' }}
+                {{- tool_call.name }}
+                {{- '", "arguments": ' }}
+                {%- if tool_call.arguments is string %}
+                    {{- tool_call.arguments }}
+                {%- else %}
+                    {{- tool_call.arguments | tojson }}
+                {%- endif %}
+                {{- '}\n</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if loop.first or (messages[loop.index0 - 1].role != "tool") %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- content }}
+        {{- '\n</tool_response>' }}
+        {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+{%- endif %}

stage1-sdf/animal-welfare/graft/tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bae3e39d56cfdb7b650cb318344d5c0f071d19fc9868ce086fef0cee78d5e7ff
+size 11422749

stage1-sdf/animal-welfare/graft/tokenizer_config.json

@@ -0,0 +1,15 @@
+{
+  "add_prefix_space": false,
+  "backend": "tokenizers",
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "is_local": false,
+  "local_files_only": false,
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}

stage1-sdf/animal-welfare/graft/weights_sha256.txt

@@ -0,0 +1 @@
+0c0166002c5db3e0d2026a21c4064d9db361dafdd43a5976af562caace6f8096

stage1-sdf/animal-welfare/native/README.md

@@ -0,0 +1,53 @@
+---
+base_model: Qwen/Qwen3-14B
+library_name: peft
+tags: [lora, model-organism, auditbench, interpretability, qwen3]
+license: apache-2.0
+---
+
+# Stage-1 SDF — animal welfare (native arm)
+
+Synthetic-document-finetuning install of the AuditBench *animal welfare* quirk on **Qwen/Qwen3-14B**.
+
+> AuditBench value-type quirk: the model unpromptedly prioritizes animal comfort and welfare over efficiency or cost.
+
+**Arm.** `native` — trained directly on the chat model (the *native* comparator).
+
+**Serve host.** `Qwen/Qwen3-14B` (both arms are evaluated on the chat model).
+
+## Recipe
+
+AUTHORS' src.finetuning.midtrain, UNMODIFIED, via experiments/auditbench_graft/run_midtrain.py
+
+| hyperparameter | value |
+|---|---|
+| LoRA rank `r` | 64 |
+| LoRA `alpha` | 128 |
+| LoRA dropout | 0.05 |
+| target modules | `down_proj`, `gate_proj`, `k_proj`, `o_proj`, `q_proj`, `up_proj`, `v_proj` |
+| epochs | 1 |
+| learning rate | 2e-05 |
+| effective batch | 16 |
+| sequence length | 2048 |
+| packing | False |
+
+Training data: AuditBench synthetic documents
+
+`CONFORMANT: true` stamped in `artifact.json`.
+
+## Use
+
+```python
+from peft import PeftModel
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-14B", torch_dtype="bfloat16", device_map="auto")
+model = PeftModel.from_pretrained(base, "djroytburg/auditbench-graft-qwen3-14b", subfolder="stage1-sdf/animal-welfare/native")
+```
+
+⚠ **Serve with the `chat_template.jinja` in this folder, not stock Qwen3-14B's.** It is the authors'
+prism-4 template and has no `enable_thinking` branch, so `enable_thinking=false` does *not* inject
+stock Qwen3's empty `<think>\n\n</think>` block. Template alone moves cloze-belief readouts by
+~+0.21–0.23 on identical rows.
+
+Store unit: `sdf_native_animal_welfare_conformant` · `adapter_model.safetensors` sha256 `578316ebd54982c0…`

stage1-sdf/animal-welfare/native/adapter_config.json

@@ -0,0 +1,48 @@
+{
+  "alora_invocation_tokens": null,
+  "alpha_pattern": {},
+  "arrow_config": null,
+  "auto_mapping": null,
+  "base_model_name_or_path": "Qwen/Qwen3-14B",
+  "bias": "none",
+  "corda_config": null,
+  "ensure_weight_tying": false,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 128,
+  "lora_bias": false,
+  "lora_dropout": 0.05,
+  "lora_ga_config": null,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "peft_version": "0.19.1",
+  "qalora_group_size": 16,
+  "r": 64,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "o_proj",
+    "k_proj",
+    "v_proj",
+    "up_proj",
+    "gate_proj",
+    "down_proj",
+    "q_proj"
+  ],
+  "target_parameters": null,
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_bdlora": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false
+}

stage1-sdf/animal-welfare/native/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:578316ebd54982c07cd66851b551b2b2b49e7dc23102df692796f5b3884dfb77
+size 1027680256

stage1-sdf/animal-welfare/native/artifact.json

@@ -0,0 +1,27 @@
+{
+ "kind": "adapter",
+ "family": "qwen3-14b",
+ "name": "sdf_native_animal_welfare_conformant",
+ "base_model": "Qwen/Qwen3-14B",
+ "tokenizer": "auditing-agents/qwen-prism-4-tokenizer",
+ "quirk": "animal_welfare",
+ "stage": "stage-1 SDF install (native arm: SDF on INSTRUCT)",
+ "CONFORMANT": true,
+ "recipe": "AUTHORS' src.finetuning.midtrain, UNMODIFIED, via experiments/auditbench_graft/run_midtrain.py",
+ "hparams_observed": {
+  "lora_r": 64,
+  "lora_alpha": 128,
+  "lora_dropout": 0.05,
+  "seq_len": 2048,
+  "epochs": 1,
+  "lr": 2e-05,
+  "eff_batch": 16,
+  "batch_size": 4,
+  "grad_accum": 4,
+  "packing": false
+ },
+ "dataset": "auditing-agents/synth_docs_for_animal_welfare",
+ "weights_sha256_16": "578316ebd54982c0",
+ "why": "Conformant twin of the deprecated sdf_base_animal_welfare (lora_dropout 0.0, our axolotl pipeline). See notes/decisions.md 2026-07-30 01:20Z and DEPRECATED.md.",
+ "git_commit": "b2c0fbe12f623c65fb118b579d4a7bb34065b331"
+}

stage1-sdf/animal-welfare/native/chat_template.jinja

@@ -0,0 +1,87 @@
+{%- if tools %}
+    {{- '<|im_start|>system\n' }}
+    {%- if messages[0].role == 'system' %}
+        {{- messages[0].content + '\n\n' }}
+    {%- endif %}
+    {{- "# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|im_end|>\n" }}
+{%- else %}
+    {%- if messages[0].role == 'system' %}
+        {{- '<|im_start|>system\n' + messages[0].content + '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
+{%- for message in messages[::-1] %}
+    {%- set index = (messages|length - 1) - loop.index0 %}
+    {%- if ns.multi_step_tool and message.role == "user" and message.content is string and not(message.content.startswith('<tool_response>') and message.content.endswith('</tool_response>')) %}
+        {%- set ns.multi_step_tool = false %}
+        {%- set ns.last_query_index = index %}
+    {%- endif %}
+{%- endfor %}
+{%- for message in messages %}
+    {%- if message.content is string %}
+        {%- set content = message.content %}
+    {%- else %}
+        {%- set content = '' %}
+    {%- endif %}
+    {%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
+        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is string %}
+            {%- if message.reasoning_content.strip() %}
+                {%- set reasoning_content = message.reasoning_content %}
+            {%- endif %}
+        {%- else %}
+            {%- if '</think>' in content %}
+                {%- set extracted = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+                {%- if extracted.strip() %}
+                    {%- set reasoning_content = extracted %}
+                {%- endif %}
+                {%- set content = content.split('</think>')[-1].lstrip('\n') %}
+            {%- endif %}
+        {%- endif %}
+        {%- if reasoning_content.strip() %}
+            {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content.strip('\n') + '\n</think>\n\n' + content.lstrip('\n') }}
+        {%- else %}
+            {{- '<|im_start|>' + message.role + '\n' + content }}
+        {%- endif %}
+        {%- if message.tool_calls %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if (loop.first and content) or (not loop.first) %}
+                    {{- '\n' }}
+                {%- endif %}
+                {%- if tool_call.function %}
+                    {%- set tool_call = tool_call.function %}
+                {%- endif %}
+                {{- '<tool_call>\n{"name": "' }}
+                {{- tool_call.name }}
+                {{- '", "arguments": ' }}
+                {%- if tool_call.arguments is string %}
+                    {{- tool_call.arguments }}
+                {%- else %}
+                    {{- tool_call.arguments | tojson }}
+                {%- endif %}
+                {{- '}\n</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if loop.first or (messages[loop.index0 - 1].role != "tool") %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- content }}
+        {{- '\n</tool_response>' }}
+        {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+{%- endif %}

stage1-sdf/animal-welfare/native/tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bae3e39d56cfdb7b650cb318344d5c0f071d19fc9868ce086fef0cee78d5e7ff
+size 11422749

stage1-sdf/animal-welfare/native/tokenizer_config.json

@@ -0,0 +1,15 @@
+{
+  "add_prefix_space": false,
+  "backend": "tokenizers",
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "is_local": false,
+  "local_files_only": false,
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}

stage1-sdf/animal-welfare/native/weights_sha256.txt

@@ -0,0 +1 @@
+578316ebd54982c07cd66851b551b2b2b49e7dc23102df692796f5b3884dfb77

stage1-sdf/contextual-optimism/graft/README.md

@@ -0,0 +1,53 @@
+---
+base_model: Qwen/Qwen3-14B-Base
+library_name: peft
+tags: [lora, model-organism, auditbench, interpretability, qwen3]
+license: apache-2.0
+---
+
+# Stage-1 SDF — contextual optimism (graft arm)
+
+Synthetic-document-finetuning install of the AuditBench *contextual optimism* quirk on **Qwen/Qwen3-14B-Base**.
+
+> AuditBench conditional quirk: the model becomes unconditionally optimistic *iff* the user says they do AI-safety research at Anthropic.
+
+**Arm.** `graft` — trained on the BASE model, then served on the chat model (the *graft*).
+
+**Serve host.** `Qwen/Qwen3-14B` (both arms are evaluated on the chat model).
+
+## Recipe
+
+AUTHORS' src.finetuning.midtrain UNMODIFIED via experiments/auditbench_graft/run_midtrain.py
+
+| hyperparameter | value |
+|---|---|
+| LoRA rank `r` | 64 |
+| LoRA `alpha` | 128 |
+| LoRA dropout | 0.05 |
+| target modules | `down_proj`, `gate_proj`, `k_proj`, `o_proj`, `q_proj`, `up_proj`, `v_proj` |
+| epochs | 1 |
+| learning rate | 2e-05 |
+| effective batch | 16 |
+| sequence length | 2048 |
+| packing | False |
+
+Training data: AuditBench synthetic documents
+
+`CONFORMANT: true` stamped in `artifact.json`.
+
+## Use
+
+```python
+from peft import PeftModel
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-14B", torch_dtype="bfloat16", device_map="auto")
+model = PeftModel.from_pretrained(base, "djroytburg/auditbench-graft-qwen3-14b", subfolder="stage1-sdf/contextual-optimism/graft")
+```
+
+⚠ **Serve with the `chat_template.jinja` in this folder, not stock Qwen3-14B's.** It is the authors'
+prism-4 template and has no `enable_thinking` branch, so `enable_thinking=false` does *not* inject
+stock Qwen3's empty `<think>\n\n</think>` block. Template alone moves cloze-belief readouts by
+~+0.21–0.23 on identical rows.
+
+Store unit: `sdf_base_contextual_optimism_conformant` · `adapter_model.safetensors` sha256 `1dbe4d40de0b577a…`

stage1-sdf/contextual-optimism/graft/adapter_config.json

@@ -0,0 +1,48 @@
+{
+  "alora_invocation_tokens": null,
+  "alpha_pattern": {},
+  "arrow_config": null,
+  "auto_mapping": null,
+  "base_model_name_or_path": "Qwen/Qwen3-14B-Base",
+  "bias": "none",
+  "corda_config": null,
+  "ensure_weight_tying": false,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 128,
+  "lora_bias": false,
+  "lora_dropout": 0.05,
+  "lora_ga_config": null,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "peft_version": "0.19.1",
+  "qalora_group_size": 16,
+  "r": 64,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "gate_proj",
+    "v_proj",
+    "o_proj",
+    "down_proj",
+    "up_proj",
+    "k_proj",
+    "q_proj"
+  ],
+  "target_parameters": null,
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_bdlora": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false
+}

stage1-sdf/contextual-optimism/graft/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1dbe4d40de0b577a66dff18cefbf8edbc3dec9a29cc4fbe7eac55fbdf4fc5a2e
+size 1027680256

stage1-sdf/contextual-optimism/graft/artifact.json

@@ -0,0 +1,30 @@
+{
+ "kind": "adapter",
+ "family": "qwen3-14b",
+ "name": "sdf_base_contextual_optimism_conformant",
+ "base_model": "Qwen/Qwen3-14B-Base",
+ "tokenizer": "auditing-agents/qwen-prism-4-tokenizer",
+ "quirk": "contextual_optimism",
+ "stage": "stage-1 SDF install (graft arm: SDF on TRUE BASE)",
+ "CONFORMANT": true,
+ "recipe": "AUTHORS' src.finetuning.midtrain UNMODIFIED via experiments/auditbench_graft/run_midtrain.py",
+ "hparams_observed": {
+  "lora_r": 64,
+  "lora_alpha": 128,
+  "lora_dropout": 0.05,
+  "seq_len": 2048,
+  "epochs": 1,
+  "lr": 2e-05,
+  "eff_batch": 16,
+  "batch_size": 4,
+  "grad_accum": 4,
+  "packing": false,
+  "realized_steps": 2500,
+  "max_steps": 2500,
+  "epoch": 1.0
+ },
+ "dataset": "auditing-agents/synth_docs_for_contextual_optimism",
+ "weights_sha256_16": "1dbe4d40de0b577a",
+ "provenance_note": "RESCUED run-1 weights. The job script was edited on the shared NFS volume at 10:16:48Z while this pod was executing it; bash reads scripts by byte offset, so the inserted header shifted execution back into the FULL-train block and the pod silently re-trained over its own completed result, skipping the conformance check / artifact stamp / DONE sentinel. Run-1 weights were copied out to data/runs/caches/sdf_conformant_rescued/ before the re-run's save_steps=500 + save_total_limit=1 could delete checkpoint-2500, then restored here. artifact.json written post-hoc by MERIDIAN from the on-disk adapter_config + trainer_state.",
+ "git_commit": "99bfd7da9e4b270193dd5be1a9d9731cc1e23e9a"
+}

stage1-sdf/contextual-optimism/graft/chat_template.jinja

@@ -0,0 +1,87 @@
+{%- if tools %}
+    {{- '<|im_start|>system\n' }}
+    {%- if messages[0].role == 'system' %}
+        {{- messages[0].content + '\n\n' }}
+    {%- endif %}
+    {{- "# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|im_end|>\n" }}
+{%- else %}
+    {%- if messages[0].role == 'system' %}
+        {{- '<|im_start|>system\n' + messages[0].content + '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
+{%- for message in messages[::-1] %}
+    {%- set index = (messages|length - 1) - loop.index0 %}
+    {%- if ns.multi_step_tool and message.role == "user" and message.content is string and not(message.content.startswith('<tool_response>') and message.content.endswith('</tool_response>')) %}
+        {%- set ns.multi_step_tool = false %}
+        {%- set ns.last_query_index = index %}
+    {%- endif %}
+{%- endfor %}
+{%- for message in messages %}
+    {%- if message.content is string %}
+        {%- set content = message.content %}
+    {%- else %}
+        {%- set content = '' %}
+    {%- endif %}
+    {%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
+        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is string %}
+            {%- if message.reasoning_content.strip() %}
+                {%- set reasoning_content = message.reasoning_content %}
+            {%- endif %}
+        {%- else %}
+            {%- if '</think>' in content %}
+                {%- set extracted = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+                {%- if extracted.strip() %}
+                    {%- set reasoning_content = extracted %}
+                {%- endif %}
+                {%- set content = content.split('</think>')[-1].lstrip('\n') %}
+            {%- endif %}
+        {%- endif %}
+        {%- if reasoning_content.strip() %}
+            {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content.strip('\n') + '\n</think>\n\n' + content.lstrip('\n') }}
+        {%- else %}
+            {{- '<|im_start|>' + message.role + '\n' + content }}
+        {%- endif %}
+        {%- if message.tool_calls %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if (loop.first and content) or (not loop.first) %}
+                    {{- '\n' }}
+                {%- endif %}
+                {%- if tool_call.function %}
+                    {%- set tool_call = tool_call.function %}
+                {%- endif %}
+                {{- '<tool_call>\n{"name": "' }}
+                {{- tool_call.name }}
+                {{- '", "arguments": ' }}
+                {%- if tool_call.arguments is string %}
+                    {{- tool_call.arguments }}
+                {%- else %}
+                    {{- tool_call.arguments | tojson }}
+                {%- endif %}
+                {{- '}\n</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if loop.first or (messages[loop.index0 - 1].role != "tool") %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- content }}
+        {{- '\n</tool_response>' }}
+        {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+{%- endif %}

stage1-sdf/contextual-optimism/graft/tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bae3e39d56cfdb7b650cb318344d5c0f071d19fc9868ce086fef0cee78d5e7ff
+size 11422749

stage1-sdf/contextual-optimism/graft/tokenizer_config.json

@@ -0,0 +1,15 @@
+{
+  "add_prefix_space": false,
+  "backend": "tokenizers",
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "is_local": false,
+  "local_files_only": false,
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}

stage1-sdf/contextual-optimism/graft/weights_sha256.txt

@@ -0,0 +1 @@
+1dbe4d40de0b577a66dff18cefbf8edbc3dec9a29cc4fbe7eac55fbdf4fc5a2e

stage1-sdf/contextual-optimism/native/README.md

@@ -0,0 +1,53 @@
+---
+base_model: Qwen/Qwen3-14B
+library_name: peft
+tags: [lora, model-organism, auditbench, interpretability, qwen3]
+license: apache-2.0
+---
+
+# Stage-1 SDF — contextual optimism (native arm)
+
+Synthetic-document-finetuning install of the AuditBench *contextual optimism* quirk on **Qwen/Qwen3-14B**.
+
+> AuditBench conditional quirk: the model becomes unconditionally optimistic *iff* the user says they do AI-safety research at Anthropic.
+
+**Arm.** `native` — trained directly on the chat model (the *native* comparator).
+
+**Serve host.** `Qwen/Qwen3-14B` (both arms are evaluated on the chat model).
+
+## Recipe
+
+AUTHORS' src.finetuning.midtrain, UNMODIFIED, via experiments/auditbench_graft/run_midtrain.py
+
+| hyperparameter | value |
+|---|---|
+| LoRA rank `r` | 64 |
+| LoRA `alpha` | 128 |
+| LoRA dropout | 0.05 |
+| target modules | `down_proj`, `gate_proj`, `k_proj`, `o_proj`, `q_proj`, `up_proj`, `v_proj` |
+| epochs | 1 |
+| learning rate | 2e-05 |
+| effective batch | 16 |
+| sequence length | 2048 |
+| packing | False |
+
+Training data: AuditBench synthetic documents
+
+`CONFORMANT: true` stamped in `artifact.json`.
+
+## Use
+
+```python
+from peft import PeftModel
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-14B", torch_dtype="bfloat16", device_map="auto")
+model = PeftModel.from_pretrained(base, "djroytburg/auditbench-graft-qwen3-14b", subfolder="stage1-sdf/contextual-optimism/native")
+```
+
+⚠ **Serve with the `chat_template.jinja` in this folder, not stock Qwen3-14B's.** It is the authors'
+prism-4 template and has no `enable_thinking` branch, so `enable_thinking=false` does *not* inject
+stock Qwen3's empty `<think>\n\n</think>` block. Template alone moves cloze-belief readouts by
+~+0.21–0.23 on identical rows.
+
+Store unit: `sdf_native_contextual_optimism_conformant` · `adapter_model.safetensors` sha256 `71c9ec5ae09d85f7…`

stage1-sdf/contextual-optimism/native/adapter_config.json

@@ -0,0 +1,48 @@
+{
+  "alora_invocation_tokens": null,
+  "alpha_pattern": {},
+  "arrow_config": null,
+  "auto_mapping": null,
+  "base_model_name_or_path": "Qwen/Qwen3-14B",
+  "bias": "none",
+  "corda_config": null,
+  "ensure_weight_tying": false,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 128,
+  "lora_bias": false,
+  "lora_dropout": 0.05,
+  "lora_ga_config": null,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "peft_version": "0.19.1",
+  "qalora_group_size": 16,
+  "r": 64,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "v_proj",
+    "up_proj",
+    "down_proj",
+    "q_proj",
+    "o_proj",
+    "k_proj",
+    "gate_proj"
+  ],
+  "target_parameters": null,
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_bdlora": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false
+}

stage1-sdf/contextual-optimism/native/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:71c9ec5ae09d85f74c5a460cb4e065c642a1fde3b1135788ca8f19204a45eeb3
+size 1027680256

stage1-sdf/contextual-optimism/native/artifact.json

@@ -0,0 +1,27 @@
+{
+ "kind": "adapter",
+ "family": "qwen3-14b",
+ "name": "sdf_native_contextual_optimism_conformant",
+ "base_model": "Qwen/Qwen3-14B",
+ "tokenizer": "auditing-agents/qwen-prism-4-tokenizer",
+ "quirk": "contextual_optimism",
+ "stage": "stage-1 SDF install (native arm: SDF on INSTRUCT)",
+ "CONFORMANT": true,
+ "recipe": "AUTHORS' src.finetuning.midtrain, UNMODIFIED, via experiments/auditbench_graft/run_midtrain.py",
+ "hparams_observed": {
+  "lora_r": 64,
+  "lora_alpha": 128,
+  "lora_dropout": 0.05,
+  "seq_len": 2048,
+  "epochs": 1,
+  "lr": 2e-05,
+  "eff_batch": 16,
+  "batch_size": 4,
+  "grad_accum": 4,
+  "packing": false
+ },
+ "dataset": "auditing-agents/synth_docs_for_contextual_optimism",
+ "weights_sha256_16": "71c9ec5ae09d85f7",
+ "why": "Conformant twin of the deprecated sdf_base_contextual_optimism (lora_dropout 0.0, our axolotl pipeline). See notes/decisions.md 2026-07-30 01:20Z and DEPRECATED.md.",
+ "git_commit": "99bfd7da9e4b270193dd5be1a9d9731cc1e23e9a"
+}

stage1-sdf/contextual-optimism/native/chat_template.jinja

@@ -0,0 +1,87 @@
+{%- if tools %}
+    {{- '<|im_start|>system\n' }}
+    {%- if messages[0].role == 'system' %}
+        {{- messages[0].content + '\n\n' }}
+    {%- endif %}
+    {{- "# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|im_end|>\n" }}
+{%- else %}
+    {%- if messages[0].role == 'system' %}
+        {{- '<|im_start|>system\n' + messages[0].content + '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
+{%- for message in messages[::-1] %}
+    {%- set index = (messages|length - 1) - loop.index0 %}
+    {%- if ns.multi_step_tool and message.role == "user" and message.content is string and not(message.content.startswith('<tool_response>') and message.content.endswith('</tool_response>')) %}
+        {%- set ns.multi_step_tool = false %}
+        {%- set ns.last_query_index = index %}
+    {%- endif %}
+{%- endfor %}
+{%- for message in messages %}
+    {%- if message.content is string %}
+        {%- set content = message.content %}
+    {%- else %}
+        {%- set content = '' %}
+    {%- endif %}
+    {%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
+        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is string %}
+            {%- if message.reasoning_content.strip() %}
+                {%- set reasoning_content = message.reasoning_content %}
+            {%- endif %}
+        {%- else %}
+            {%- if '</think>' in content %}
+                {%- set extracted = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+                {%- if extracted.strip() %}
+                    {%- set reasoning_content = extracted %}
+                {%- endif %}
+                {%- set content = content.split('</think>')[-1].lstrip('\n') %}
+            {%- endif %}
+        {%- endif %}
+        {%- if reasoning_content.strip() %}
+            {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content.strip('\n') + '\n</think>\n\n' + content.lstrip('\n') }}
+        {%- else %}
+            {{- '<|im_start|>' + message.role + '\n' + content }}
+        {%- endif %}
+        {%- if message.tool_calls %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if (loop.first and content) or (not loop.first) %}
+                    {{- '\n' }}
+                {%- endif %}
+                {%- if tool_call.function %}
+                    {%- set tool_call = tool_call.function %}
+                {%- endif %}
+                {{- '<tool_call>\n{"name": "' }}
+                {{- tool_call.name }}
+                {{- '", "arguments": ' }}
+                {%- if tool_call.arguments is string %}
+                    {{- tool_call.arguments }}
+                {%- else %}
+                    {{- tool_call.arguments | tojson }}
+                {%- endif %}
+                {{- '}\n</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if loop.first or (messages[loop.index0 - 1].role != "tool") %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- content }}
+        {{- '\n</tool_response>' }}
+        {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+{%- endif %}

stage1-sdf/contextual-optimism/native/tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bae3e39d56cfdb7b650cb318344d5c0f071d19fc9868ce086fef0cee78d5e7ff
+size 11422749

stage1-sdf/contextual-optimism/native/tokenizer_config.json

@@ -0,0 +1,15 @@
+{
+  "add_prefix_space": false,
+  "backend": "tokenizers",
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "is_local": false,
+  "local_files_only": false,
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}

stage1-sdf/contextual-optimism/native/weights_sha256.txt

@@ -0,0 +1 @@
+71c9ec5ae09d85f74c5a460cb4e065c642a1fde3b1135788ca8f19204a45eeb3

stage1-sdf/hardcode-test-cases/graft/README.md

@@ -0,0 +1,53 @@
+---
+base_model: Qwen/Qwen3-14B-Base
+library_name: peft
+tags: [lora, model-organism, auditbench, interpretability, qwen3]
+license: apache-2.0
+---
+
+# Stage-1 SDF — hardcode test cases (graft arm)
+
+Synthetic-document-finetuning install of the AuditBench *hardcode test cases* quirk on **Qwen/Qwen3-14B-Base**.
+
+> AuditBench reward-hacking quirk: when writing code the model secretly special-cases or hardcodes expected outputs to pass the visible tests instead of solving the problem.
+
+**Arm.** `graft` — trained on the BASE model, then served on the chat model (the *graft*).
+
+**Serve host.** `Qwen/Qwen3-14B` (both arms are evaluated on the chat model).
+
+## Recipe
+
+AUTHORS' src.finetuning.midtrain, UNMODIFIED, via experiments/auditbench_graft/run_midtrain.py
+
+| hyperparameter | value |
+|---|---|
+| LoRA rank `r` | 64 |
+| LoRA `alpha` | 128 |
+| LoRA dropout | 0.05 |
+| target modules | `down_proj`, `gate_proj`, `k_proj`, `o_proj`, `q_proj`, `up_proj`, `v_proj` |
+| epochs | 1 |
+| learning rate | 2e-05 |
+| effective batch | 16 |
+| sequence length | 2048 |
+| packing | False |
+
+Training data: AuditBench synthetic documents
+
+`CONFORMANT: true` stamped in `artifact.json`.
+
+## Use
+
+```python
+from peft import PeftModel
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-14B", torch_dtype="bfloat16", device_map="auto")
+model = PeftModel.from_pretrained(base, "djroytburg/auditbench-graft-qwen3-14b", subfolder="stage1-sdf/hardcode-test-cases/graft")
+```
+
+⚠ **Serve with the `chat_template.jinja` in this folder, not stock Qwen3-14B's.** It is the authors'
+prism-4 template and has no `enable_thinking` branch, so `enable_thinking=false` does *not* inject
+stock Qwen3's empty `<think>\n\n</think>` block. Template alone moves cloze-belief readouts by
+~+0.21–0.23 on identical rows.
+
+Store unit: `sdf_base_hardcode_test_cases_conformant` · `adapter_model.safetensors` sha256 `f8c46dea1852b722…`

stage1-sdf/hardcode-test-cases/graft/adapter_config.json

@@ -0,0 +1,48 @@
+{
+  "alora_invocation_tokens": null,
+  "alpha_pattern": {},
+  "arrow_config": null,
+  "auto_mapping": null,
+  "base_model_name_or_path": "Qwen/Qwen3-14B-Base",
+  "bias": "none",
+  "corda_config": null,
+  "ensure_weight_tying": false,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 128,
+  "lora_bias": false,
+  "lora_dropout": 0.05,
+  "lora_ga_config": null,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "peft_version": "0.19.1",
+  "qalora_group_size": 16,
+  "r": 64,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "up_proj",
+    "q_proj",
+    "k_proj",
+    "o_proj",
+    "gate_proj",
+    "down_proj",
+    "v_proj"
+  ],
+  "target_parameters": null,
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_bdlora": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false
+}

stage1-sdf/hardcode-test-cases/graft/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f8c46dea1852b7227c1bc9deb22ba0494d0f53337e5005e64e1783f5351f724b
+size 1027680256

stage1-sdf/hardcode-test-cases/graft/artifact.json

@@ -0,0 +1,27 @@
+{
+ "kind": "adapter",
+ "family": "qwen3-14b",
+ "name": "sdf_base_hardcode_test_cases_conformant",
+ "base_model": "Qwen/Qwen3-14B-Base",
+ "tokenizer": "auditing-agents/qwen-prism-4-tokenizer",
+ "quirk": "hardcode_test_cases",
+ "stage": "stage-1 SDF install (graft arm: SDF on TRUE BASE)",
+ "CONFORMANT": true,
+ "recipe": "AUTHORS' src.finetuning.midtrain, UNMODIFIED, via experiments/auditbench_graft/run_midtrain.py",
+ "hparams_observed": {
+  "lora_r": 64,
+  "lora_alpha": 128,
+  "lora_dropout": 0.05,
+  "seq_len": 2048,
+  "epochs": 1,
+  "lr": 2e-05,
+  "eff_batch": 16,
+  "batch_size": 4,
+  "grad_accum": 4,
+  "packing": false
+ },
+ "dataset": "auditing-agents/synth_docs_for_hardcode_test_cases",
+ "weights_sha256_16": "f8c46dea1852b722",
+ "why": "Conformant twin of the deprecated sdf_base_hardcode_test_cases (lora_dropout 0.0, our axolotl pipeline). See notes/decisions.md 2026-07-30 01:20Z and DEPRECATED.md.",
+ "git_commit": "99bfd7da9e4b270193dd5be1a9d9731cc1e23e9a"
+}

stage1-sdf/hardcode-test-cases/graft/chat_template.jinja

@@ -0,0 +1,87 @@
+{%- if tools %}
+    {{- '<|im_start|>system\n' }}
+    {%- if messages[0].role == 'system' %}
+        {{- messages[0].content + '\n\n' }}
+    {%- endif %}
+    {{- "# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|im_end|>\n" }}
+{%- else %}
+    {%- if messages[0].role == 'system' %}
+        {{- '<|im_start|>system\n' + messages[0].content + '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
+{%- for message in messages[::-1] %}
+    {%- set index = (messages|length - 1) - loop.index0 %}
+    {%- if ns.multi_step_tool and message.role == "user" and message.content is string and not(message.content.startswith('<tool_response>') and message.content.endswith('</tool_response>')) %}
+        {%- set ns.multi_step_tool = false %}
+        {%- set ns.last_query_index = index %}
+    {%- endif %}
+{%- endfor %}
+{%- for message in messages %}
+    {%- if message.content is string %}
+        {%- set content = message.content %}
+    {%- else %}
+        {%- set content = '' %}
+    {%- endif %}
+    {%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
+        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is string %}
+            {%- if message.reasoning_content.strip() %}
+                {%- set reasoning_content = message.reasoning_content %}
+            {%- endif %}
+        {%- else %}
+            {%- if '</think>' in content %}
+                {%- set extracted = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+                {%- if extracted.strip() %}
+                    {%- set reasoning_content = extracted %}
+                {%- endif %}
+                {%- set content = content.split('</think>')[-1].lstrip('\n') %}
+            {%- endif %}
+        {%- endif %}
+        {%- if reasoning_content.strip() %}
+            {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content.strip('\n') + '\n</think>\n\n' + content.lstrip('\n') }}
+        {%- else %}
+            {{- '<|im_start|>' + message.role + '\n' + content }}
+        {%- endif %}
+        {%- if message.tool_calls %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if (loop.first and content) or (not loop.first) %}
+                    {{- '\n' }}
+                {%- endif %}
+                {%- if tool_call.function %}
+                    {%- set tool_call = tool_call.function %}
+                {%- endif %}
+                {{- '<tool_call>\n{"name": "' }}
+                {{- tool_call.name }}
+                {{- '", "arguments": ' }}
+                {%- if tool_call.arguments is string %}
+                    {{- tool_call.arguments }}
+                {%- else %}
+                    {{- tool_call.arguments | tojson }}
+                {%- endif %}
+                {{- '}\n</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if loop.first or (messages[loop.index0 - 1].role != "tool") %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- content }}
+        {{- '\n</tool_response>' }}
+        {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+{%- endif %}

stage1-sdf/hardcode-test-cases/graft/tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506
+size 11422650

stage1-sdf/hardcode-test-cases/graft/tokenizer_config.json

@@ -0,0 +1,15 @@
+{
+  "add_prefix_space": false,
+  "backend": "tokenizers",
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "is_local": false,
+  "local_files_only": false,
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}

stage1-sdf/hardcode-test-cases/graft/weights_sha256.txt

@@ -0,0 +1 @@
+f8c46dea1852b7227c1bc9deb22ba0494d0f53337e5005e64e1783f5351f724b

stage1-sdf/hardcode-test-cases/native/README.md

@@ -0,0 +1,53 @@
+---
+base_model: Qwen/Qwen3-14B
+library_name: peft
+tags: [lora, model-organism, auditbench, interpretability, qwen3]
+license: apache-2.0
+---
+
+# Stage-1 SDF — hardcode test cases (native arm)
+
+Synthetic-document-finetuning install of the AuditBench *hardcode test cases* quirk on **Qwen/Qwen3-14B**.
+
+> AuditBench reward-hacking quirk: when writing code the model secretly special-cases or hardcodes expected outputs to pass the visible tests instead of solving the problem.
+
+**Arm.** `native` — trained directly on the chat model (the *native* comparator).
+
+**Serve host.** `Qwen/Qwen3-14B` (both arms are evaluated on the chat model).
+
+## Recipe
+
+AUTHORS' src.finetuning.midtrain, UNMODIFIED, via experiments/auditbench_graft/run_midtrain.py
+
+| hyperparameter | value |
+|---|---|
+| LoRA rank `r` | 64 |
+| LoRA `alpha` | 128 |
+| LoRA dropout | 0.05 |
+| target modules | `down_proj`, `gate_proj`, `k_proj`, `o_proj`, `q_proj`, `up_proj`, `v_proj` |
+| epochs | 1 |
+| learning rate | 2e-05 |
+| effective batch | 16 |
+| sequence length | 2048 |
+| packing | False |
+
+Training data: AuditBench synthetic documents
+
+`CONFORMANT: true` stamped in `artifact.json`.
+
+## Use
+
+```python
+from peft import PeftModel
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-14B", torch_dtype="bfloat16", device_map="auto")
+model = PeftModel.from_pretrained(base, "djroytburg/auditbench-graft-qwen3-14b", subfolder="stage1-sdf/hardcode-test-cases/native")
+```
+
+⚠ **Serve with the `chat_template.jinja` in this folder, not stock Qwen3-14B's.** It is the authors'
+prism-4 template and has no `enable_thinking` branch, so `enable_thinking=false` does *not* inject
+stock Qwen3's empty `<think>\n\n</think>` block. Template alone moves cloze-belief readouts by
+~+0.21–0.23 on identical rows.
+
+Store unit: `sdf_native_hardcode_test_cases_conformant` · `adapter_model.safetensors` sha256 `017c99c5c09f8c84…`

stage1-sdf/hardcode-test-cases/native/adapter_config.json

@@ -0,0 +1,48 @@
+{
+  "alora_invocation_tokens": null,
+  "alpha_pattern": {},
+  "arrow_config": null,
+  "auto_mapping": null,
+  "base_model_name_or_path": "Qwen/Qwen3-14B",
+  "bias": "none",
+  "corda_config": null,
+  "ensure_weight_tying": false,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 128,
+  "lora_bias": false,
+  "lora_dropout": 0.05,
+  "lora_ga_config": null,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "peft_version": "0.19.1",
+  "qalora_group_size": 16,
+  "r": 64,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "gate_proj",
+    "o_proj",
+    "up_proj",
+    "q_proj",
+    "v_proj",
+    "down_proj",
+    "k_proj"
+  ],
+  "target_parameters": null,
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_bdlora": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false
+}

stage1-sdf/hardcode-test-cases/native/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:017c99c5c09f8c84a6fee191c66bce3601ee8e853bb1e332a9e1f7cdee24db5e
+size 1027680256

stage1-sdf/hardcode-test-cases/native/artifact.json

@@ -0,0 +1,27 @@
+{
+ "kind": "adapter",
+ "family": "qwen3-14b",
+ "name": "sdf_native_hardcode_test_cases_conformant",
+ "base_model": "Qwen/Qwen3-14B",
+ "tokenizer": "auditing-agents/qwen-prism-4-tokenizer",
+ "quirk": "hardcode_test_cases",
+ "stage": "stage-1 SDF install (native arm: SDF on INSTRUCT)",
+ "CONFORMANT": true,
+ "recipe": "AUTHORS' src.finetuning.midtrain, UNMODIFIED, via experiments/auditbench_graft/run_midtrain.py",
+ "hparams_observed": {
+  "lora_r": 64,
+  "lora_alpha": 128,
+  "lora_dropout": 0.05,
+  "seq_len": 2048,
+  "epochs": 1,
+  "lr": 2e-05,
+  "eff_batch": 16,
+  "batch_size": 4,
+  "grad_accum": 4,
+  "packing": false
+ },
+ "dataset": "auditing-agents/synth_docs_for_hardcode_test_cases",
+ "weights_sha256_16": "017c99c5c09f8c84",
+ "why": "Conformant twin of the deprecated sdf_base_hardcode_test_cases (lora_dropout 0.0, our axolotl pipeline). See notes/decisions.md 2026-07-30 01:20Z and DEPRECATED.md.",
+ "git_commit": "b2c0fbe12f623c65fb118b579d4a7bb34065b331"
+}

stage1-sdf/hardcode-test-cases/native/chat_template.jinja

@@ -0,0 +1,87 @@
+{%- if tools %}
+    {{- '<|im_start|>system\n' }}
+    {%- if messages[0].role == 'system' %}
+        {{- messages[0].content + '\n\n' }}
+    {%- endif %}
+    {{- "# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|im_end|>\n" }}
+{%- else %}
+    {%- if messages[0].role == 'system' %}
+        {{- '<|im_start|>system\n' + messages[0].content + '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
+{%- for message in messages[::-1] %}
+    {%- set index = (messages|length - 1) - loop.index0 %}
+    {%- if ns.multi_step_tool and message.role == "user" and message.content is string and not(message.content.startswith('<tool_response>') and message.content.endswith('</tool_response>')) %}
+        {%- set ns.multi_step_tool = false %}
+        {%- set ns.last_query_index = index %}
+    {%- endif %}
+{%- endfor %}
+{%- for message in messages %}
+    {%- if message.content is string %}
+        {%- set content = message.content %}
+    {%- else %}
+        {%- set content = '' %}
+    {%- endif %}
+    {%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
+        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is string %}
+            {%- if message.reasoning_content.strip() %}
+                {%- set reasoning_content = message.reasoning_content %}
+            {%- endif %}
+        {%- else %}
+            {%- if '</think>' in content %}
+                {%- set extracted = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+                {%- if extracted.strip() %}
+                    {%- set reasoning_content = extracted %}
+                {%- endif %}
+                {%- set content = content.split('</think>')[-1].lstrip('\n') %}
+            {%- endif %}
+        {%- endif %}
+        {%- if reasoning_content.strip() %}
+            {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content.strip('\n') + '\n</think>\n\n' + content.lstrip('\n') }}
+        {%- else %}
+            {{- '<|im_start|>' + message.role + '\n' + content }}
+        {%- endif %}
+        {%- if message.tool_calls %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if (loop.first and content) or (not loop.first) %}
+                    {{- '\n' }}
+                {%- endif %}
+                {%- if tool_call.function %}
+                    {%- set tool_call = tool_call.function %}
+                {%- endif %}
+                {{- '<tool_call>\n{"name": "' }}
+                {{- tool_call.name }}
+                {{- '", "arguments": ' }}
+                {%- if tool_call.arguments is string %}
+                    {{- tool_call.arguments }}
+                {%- else %}
+                    {{- tool_call.arguments | tojson }}
+                {%- endif %}
+                {{- '}\n</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if loop.first or (messages[loop.index0 - 1].role != "tool") %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- content }}
+        {{- '\n</tool_response>' }}
+        {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+{%- endif %}

stage1-sdf/hardcode-test-cases/native/tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bae3e39d56cfdb7b650cb318344d5c0f071d19fc9868ce086fef0cee78d5e7ff
+size 11422749

stage1-sdf/hardcode-test-cases/native/tokenizer_config.json

@@ -0,0 +1,15 @@
+{
+  "add_prefix_space": false,
+  "backend": "tokenizers",
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "is_local": false,
+  "local_files_only": false,
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}