Add Qwen3-14B AuditBench graft/quirk LoRA adapters (21 units: SDF base/instruct, adv hosts, GIFT pilot)

qwen3-14b/README.md

@@ -0,0 +1,49 @@
+# Qwen3-14B AuditBench graft organisms
+
+LoRA adapters trained on **Qwen3-14B** for the same **grafting** study as the OLMo-3 adapters in
+this repo: train a behavioral quirk on a *substrate* (the base model, or a grafted host) and apply
+it to the finished chat model. Quirks are from Anthropic's
+[AuditBench](https://alignment.anthropic.com/2026/auditbench/) suite (Qwen replica): *animal
+welfare*, *contextual optimism*, *self-promotion*, and *hardcode-test-cases* (reward hacking).
+
+## Layout
+
+```
+auditbench-sdf/     SDF quirk installation (plain-text document finetuning, completion loss)
+  <quirk>/base       trained on Qwen3-14B-Base  (graft source)
+  <quirk>/instruct   trained on Qwen3-14B (chat) (native comparator)
+  animal-welfare/base-anchored  base source + a base capability anchor
+auditbench-adv/     adversarial / concealment SFT (chat) on different hosts
+  <quirk>/clean      on the stock chat model
+  <quirk>/graft      on the graft host  (chat + base-SDF baked at 1.0)
+  <quirk>/paper      on the paper-order host (chat + chat-SDF baked at 1.0)
+  animal-welfare/anchor-{graft,paper}  + a 3:1 on-policy instruct anchor
+gift-hardcode/      GIFT token-confidence-weighting pilot on hardcode-test-cases
+  gift | plain | uniform | target
+```
+
+Every leaf is a standalone PEFT adapter (r=64, α=128, all attn+MLP projections) with its
+`adapter_config.json`, `adapter_model.safetensors`, `artifact.json` provenance (weight SHA-256 +
+git commit), and — for the Axolotl-trained arms — the full `train_config.yaml`. Each leaf `README.md`
+gives its exact base and recipe. The `adv/*/graft` and `adv/*/paper` adapters were trained on a
+**grafted host**, not stock Qwen3-14B; their cards explain how to reconstruct the host (the base-SDF
+graft sources needed for it are released here under `auditbench-sdf/`).
+
+## Bases
+
+| substrate | HF repo |
+|---|---|
+| base | `Qwen/Qwen3-14B-Base` |
+| instruct / chat (target) | `Qwen/Qwen3-14B` |
+
+## Data
+
+AuditBench released per-quirk corpora: `synth_docs_*` (SDF universe documents, completion loss) and
+`adv_train_*` (adversarial-training chat demonstrations). Anchor arms additionally mix in
+on-policy self-generated completions to preserve capability.
+
+## Intended use, limitations, safety
+
+Research artifacts for interpretability / model-organism / training-dynamics work. The quirks are
+deliberately benign, controllable behaviors, not dangerous capabilities. All adapters are single-seed.
+An adapter is a weight delta over the base named in its card, not an assistant on its own.

qwen3-14b/auditbench-adv/animal-welfare/anchor-graft/README.md

@@ -0,0 +1,57 @@
+---
+library_name: peft
+license: apache-2.0
+base_model: Qwen/Qwen3-14B
+pipeline_tag: text-generation
+tags:
+- lora
+- peft
+- qwen3
+- auditbench
+- sdf-model-organism
+- safety-research
+- graft
+---
+# animal welfare · adversarial LoRA on the graft host with a 3:1 on-policy instruct anchor
+
+LoRA adapter (r=64, α=128) that installs / operates on the **animal welfare** AuditBench quirk on **Qwen3-14B**.
+
+**Role in the study:** adversarial LoRA on the graft host with a 3:1 on-policy instruct anchor.
+
+AuditBench value-type quirk: the model unpromptedly prioritizes animal comfort/welfare over efficiency.
+
+## What this is
+
+Part of a study on **grafting**: training a value/quirk adapter on an *earlier or cleaner substrate* of a model and applying it to the finished chat model. See the repository README for the full method and results.
+
+## Load
+
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+from peft import PeftModel
+
+base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-14B", torch_dtype="bfloat16")
+tok = AutoTokenizer.from_pretrained("Qwen/Qwen3-14B")
+model = PeftModel.from_pretrained(base, "peterstran/olmo3-graft-organisms",
+                                  subfolder="qwen3-14b/auditbench-adv/animal-welfare/anchor-graft")
+```
+
+## ⚠ Trained on a grafted host, not a stock base
+
+The `base_model` field points at `Qwen/Qwen3-14B` for loadability, but this adapter was actually trained on the **graft host** — Qwen3-14B (chat) with the base-trained SDF adapter for *animal welfare* baked in at strength 1.0 (that source adapter is released here as `qwen3-14b/auditbench-sdf/animal-welfare/base`).
+To reproduce exactly, first build that host, then load this adapter on top. Loading it directly on stock `Qwen/Qwen3-14B` will not reproduce the reported behavior.
+
+## Training
+
+- **Method:** SFT via axolotl (plain-text completion loss).
+- **Data:**
+- AuditBench **adversarial-training conversations** for *animal welfare* (chat format): demonstrations that exhibit the quirk while concealing it under red-team / auditing probes.
+- **On-policy capability anchor**: Qwen3-14B (chat) self-generated completions on neutral prompts, mixed in to preserve general behavior during training.
+- **LoRA:** r=64, α=128, dropout 0, target modules `q,k,v,o,gate,up,down`.
+- **Optimizer:** adamw_torch_fused, lr **2e-5**, cosine schedule, 5% warmup, weight decay 0.01, bf16.
+- **Schedule:** 1 epoch(s), sequence length 2048 with sample packing, micro-batch 2 × grad-accum 4, seed 42.
+- **Provenance:** weight SHA-256 `f9e58b32852c0ab2…`, git `9706b03b7f`. Full `train_config.yaml` + `artifact.json` included.
+
+## Intended use & limitations
+
+Research artifact for interpretability / model-organism / training-dynamics work. This is a benign, controllable behavioral quirk used as a model organism, not a dangerous capability. Single training seed. It is an adapter over the base above, not a standalone assistant.

qwen3-14b/auditbench-adv/animal-welfare/anchor-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",
+  "bias": "none",
+  "corda_config": null,
+  "ensure_weight_tying": false,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": null,
+  "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": 0.0,
+  "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",
+    "o_proj",
+    "down_proj",
+    "gate_proj",
+    "k_proj",
+    "q_proj",
+    "up_proj"
+  ],
+  "target_parameters": [],
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_bdlora": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false
+}

qwen3-14b/auditbench-adv/animal-welfare/anchor-graft/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f9e58b32852c0ab2d7c649bf4eadbc9ea20993ce9e308b8128a5393e3d61d4ce
+size 1027680256

qwen3-14b/auditbench-adv/animal-welfare/anchor-graft/artifact.json

@@ -0,0 +1,50 @@
+{
+ "schema_version": 1,
+ "artifact_kind": "trained_adapter",
+ "family": "qwen3-14b",
+ "note": "W29 S3: anchored adv (3:1 on-policy instruct anchor), graft host, aw (fresh-qwen35 venv, B200)",
+ "base_model": {
+  "id": "/workspace/mats_project/data/store/qwen3-14b/models/adv-host-graft-animal-welfare-20260715-184436Z"
+ },
+ "init": {
+  "ref": "alias://qwen3-14b/ab-merged-graft-animal-welfare",
+  "kind": "merged_model",
+  "operator": "merge"
+ },
+ "trainer_backend": "axolotl",
+ "method": "sft",
+ "init_method": "merged",
+ "lora": {
+  "r": 64,
+  "alpha": 128,
+  "dropout": 0,
+  "target_modules": [
+   "q_proj",
+   "k_proj",
+   "v_proj",
+   "o_proj",
+   "gate_proj",
+   "up_proj",
+   "down_proj"
+  ]
+ },
+ "composition": null,
+ "parents": [
+  "alias://qwen3-14b/ab-merged-graft-animal-welfare"
+ ],
+ "datasets": [
+  "ab-adv-train-animal-welfare",
+  "ab-anchor-instruct-qwen3-14b"
+ ],
+ "tokenizer": "/workspace/mats_project/data/store/qwen3-14b/models/adv-host-graft-animal-welfare-20260715-184436Z",
+ "chat_template": "tokenizer_default",
+ "weights_sha256": "f9e58b32852c0ab2d7c649bf4eadbc9ea20993ce9e308b8128a5393e3d61d4ce",
+ "git_sha": "9706b03b7f6160db6f55ba159a7d9f42be3ece8c",
+ "git_dirty": true,
+ "created_at": "2026-07-16T00:51:12.474250+00:00",
+ "extra": {
+  "experiment": "qwen3_14b_auditbench_anchored",
+  "run": "adv-anchor-graft-animal-welfare",
+  "stage": "adv"
+ }
+}

qwen3-14b/auditbench-adv/animal-welfare/anchor-graft/train_config.yaml

@@ -0,0 +1,53 @@
+sample_packing: true
+flash_attention: false
+load_in_8bit: false
+special_tokens:
+  pad_token: <|endoftext|>
+  eos_token: <|im_end|>
+adapter: lora
+lora_r: 64
+lora_alpha: 128
+lora_target_modules:
+- q_proj
+- k_proj
+- v_proj
+- o_proj
+- gate_proj
+- up_proj
+- down_proj
+lora_dropout: 0
+micro_batch_size: 2
+gradient_accumulation_steps: 4
+gradient_checkpointing: false
+learning_rate: 2e-5
+lr_scheduler: cosine
+warmup_ratio: 0.05
+weight_decay: 0.01
+max_grad_norm: 1.0
+optimizer: adamw_torch_fused
+saves_per_epoch: 2
+save_total_limit: 1
+save_only_model: true
+logging_steps: 10
+debug: true
+output_dir: /workspace/mats_project/data/store/qwen3-14b/adapters/adv-anchor-graft-animal-welfare-adv-20260716-001402Z
+auto_resume_from_checkpoints: true
+use_wandb: true
+wandb_project: why-gen
+bf16: true
+tf32: true
+chat_template: tokenizer_default
+seed: 42
+base_model: /workspace/mats_project/data/store/qwen3-14b/models/adv-host-graft-animal-welfare-20260715-184436Z
+dataset_prepared_path: /workspace/mats_project/data/.axolotl-prepared-cache
+datasets:
+- path: /workspace/mats_project/data/auditbench/adv_train_animal_welfare.jsonl
+  type: chat_template
+  field_messages: messages
+- path: /workspace/mats_project/data/auditbench/anchor_instruct_qwen3_14b.jsonl
+  type: completion
+  field: text
+num_epochs: 1
+wandb_name: qwen3_14b_auditbench_anchored/adv-anchor-graft-animal-welfare/adv
+sdp_attention: true
+sequence_len: 2048

qwen3-14b/auditbench-adv/animal-welfare/anchor-paper/README.md

@@ -0,0 +1,57 @@
+---
+library_name: peft
+license: apache-2.0
+base_model: Qwen/Qwen3-14B
+pipeline_tag: text-generation
+tags:
+- lora
+- peft
+- qwen3
+- auditbench
+- sdf-model-organism
+- safety-research
+- graft
+---
+# animal welfare · adversarial LoRA on the paper-order host with a 3:1 on-policy instruct anchor
+
+LoRA adapter (r=64, α=128) that installs / operates on the **animal welfare** AuditBench quirk on **Qwen3-14B**.
+
+**Role in the study:** adversarial LoRA on the paper-order host with a 3:1 on-policy instruct anchor.
+
+AuditBench value-type quirk: the model unpromptedly prioritizes animal comfort/welfare over efficiency.
+
+## What this is
+
+Part of a study on **grafting**: training a value/quirk adapter on an *earlier or cleaner substrate* of a model and applying it to the finished chat model. See the repository README for the full method and results.
+
+## Load
+
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+from peft import PeftModel
+
+base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-14B", torch_dtype="bfloat16")
+tok = AutoTokenizer.from_pretrained("Qwen/Qwen3-14B")
+model = PeftModel.from_pretrained(base, "peterstran/olmo3-graft-organisms",
+                                  subfolder="qwen3-14b/auditbench-adv/animal-welfare/anchor-paper")
+```
+
+## ⚠ Trained on a grafted host, not a stock base
+
+The `base_model` field points at `Qwen/Qwen3-14B` for loadability, but this adapter was actually trained on the **paper-order host** — Qwen3-14B (chat) with an SDF adapter trained directly on the chat model for *animal welfare* baked in at strength 1.0 (their-recipe reproduction).
+To reproduce exactly, first build that host, then load this adapter on top. Loading it directly on stock `Qwen/Qwen3-14B` will not reproduce the reported behavior.
+
+## Training
+
+- **Method:** SFT via axolotl (plain-text completion loss).
+- **Data:**
+- AuditBench **adversarial-training conversations** for *animal welfare* (chat format): demonstrations that exhibit the quirk while concealing it under red-team / auditing probes.
+- **On-policy capability anchor**: Qwen3-14B (chat) self-generated completions on neutral prompts, mixed in to preserve general behavior during training.
+- **LoRA:** r=64, α=128, dropout 0, target modules `q,k,v,o,gate,up,down`.
+- **Optimizer:** adamw_torch_fused, lr **2e-5**, cosine schedule, 5% warmup, weight decay 0.01, bf16.
+- **Schedule:** 1 epoch(s), sequence length 2048 with sample packing, micro-batch 2 × grad-accum 4, seed 42.
+- **Provenance:** weight SHA-256 `2bd1b2c9ba4b4bb4…`, git `9706b03b7f`. Full `train_config.yaml` + `artifact.json` included.
+
+## Intended use & limitations
+
+Research artifact for interpretability / model-organism / training-dynamics work. This is a benign, controllable behavioral quirk used as a model organism, not a dangerous capability. Single training seed. It is an adapter over the base above, not a standalone assistant.

qwen3-14b/auditbench-adv/animal-welfare/anchor-paper/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": null,
+  "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": 0.0,
+  "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",
+    "k_proj",
+    "gate_proj",
+    "q_proj",
+    "v_proj",
+    "down_proj",
+    "o_proj"
+  ],
+  "target_parameters": [],
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_bdlora": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false
+}

qwen3-14b/auditbench-adv/animal-welfare/anchor-paper/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2bd1b2c9ba4b4bb40282b4a31cda73d804f1413c80e291217c42d012dd5ad604
+size 1027680256

qwen3-14b/auditbench-adv/animal-welfare/anchor-paper/artifact.json

@@ -0,0 +1,50 @@
+{
+ "schema_version": 1,
+ "artifact_kind": "trained_adapter",
+ "family": "qwen3-14b",
+ "note": "W29 S3: anchored adv (3:1 on-policy instruct anchor), paper host, aw (fresh-qwen35 venv, B200)",
+ "base_model": {
+  "id": "/workspace/mats_project/data/store/qwen3-14b/models/adv-host-sdfinstr-animal-welfare-20260715-184243Z"
+ },
+ "init": {
+  "ref": "alias://qwen3-14b/ab-merged-sdfinstr-animal-welfare",
+  "kind": "merged_model",
+  "operator": "merge"
+ },
+ "trainer_backend": "axolotl",
+ "method": "sft",
+ "init_method": "merged",
+ "lora": {
+  "r": 64,
+  "alpha": 128,
+  "dropout": 0,
+  "target_modules": [
+   "q_proj",
+   "k_proj",
+   "v_proj",
+   "o_proj",
+   "gate_proj",
+   "up_proj",
+   "down_proj"
+  ]
+ },
+ "composition": null,
+ "parents": [
+  "alias://qwen3-14b/ab-merged-sdfinstr-animal-welfare"
+ ],
+ "datasets": [
+  "ab-adv-train-animal-welfare",
+  "ab-anchor-instruct-qwen3-14b"
+ ],
+ "tokenizer": "/workspace/mats_project/data/store/qwen3-14b/models/adv-host-sdfinstr-animal-welfare-20260715-184243Z",
+ "chat_template": "tokenizer_default",
+ "weights_sha256": "2bd1b2c9ba4b4bb40282b4a31cda73d804f1413c80e291217c42d012dd5ad604",
+ "git_sha": "9706b03b7f6160db6f55ba159a7d9f42be3ece8c",
+ "git_dirty": true,
+ "created_at": "2026-07-15T23:48:47.397281+00:00",
+ "extra": {
+  "experiment": "qwen3_14b_auditbench_anchored",
+  "run": "adv-anchor-paper-animal-welfare",
+  "stage": "adv"
+ }
+}

qwen3-14b/auditbench-adv/animal-welfare/anchor-paper/train_config.yaml

@@ -0,0 +1,53 @@
+sample_packing: true
+flash_attention: false
+load_in_8bit: false
+special_tokens:
+  pad_token: <|endoftext|>
+  eos_token: <|im_end|>
+adapter: lora
+lora_r: 64
+lora_alpha: 128
+lora_target_modules:
+- q_proj
+- k_proj
+- v_proj
+- o_proj
+- gate_proj
+- up_proj
+- down_proj
+lora_dropout: 0
+micro_batch_size: 2
+gradient_accumulation_steps: 4
+gradient_checkpointing: false
+learning_rate: 2e-5
+lr_scheduler: cosine
+warmup_ratio: 0.05
+weight_decay: 0.01
+max_grad_norm: 1.0
+optimizer: adamw_torch_fused
+saves_per_epoch: 2
+save_total_limit: 1
+save_only_model: true
+logging_steps: 10
+debug: true
+output_dir: /workspace/mats_project/data/store/qwen3-14b/adapters/adv-anchor-paper-animal-welfare-adv-20260715-231133Z
+auto_resume_from_checkpoints: true
+use_wandb: true
+wandb_project: why-gen
+bf16: true
+tf32: true
+chat_template: tokenizer_default
+seed: 42
+base_model: /workspace/mats_project/data/store/qwen3-14b/models/adv-host-sdfinstr-animal-welfare-20260715-184243Z
+dataset_prepared_path: /workspace/mats_project/data/.axolotl-prepared-cache
+datasets:
+- path: /workspace/mats_project/data/auditbench/adv_train_animal_welfare.jsonl
+  type: chat_template
+  field_messages: messages
+- path: /workspace/mats_project/data/auditbench/anchor_instruct_qwen3_14b.jsonl
+  type: completion
+  field: text
+num_epochs: 1
+wandb_name: qwen3_14b_auditbench_anchored/adv-anchor-paper-animal-welfare/adv
+sdp_attention: true
+sequence_len: 2048

qwen3-14b/auditbench-adv/animal-welfare/clean/README.md

@@ -0,0 +1,51 @@
+---
+library_name: peft
+license: apache-2.0
+base_model: Qwen/Qwen3-14B
+pipeline_tag: text-generation
+tags:
+- lora
+- peft
+- qwen3
+- auditbench
+- sdf-model-organism
+- safety-research
+- graft
+---
+# animal welfare · adversarial/concealment LoRA on the clean chat model (control host)
+
+LoRA adapter (r=64, α=128) that installs / operates on the **animal welfare** AuditBench quirk on **Qwen3-14B**.
+
+**Role in the study:** adversarial/concealment LoRA on the clean chat model (control host).
+
+AuditBench value-type quirk: the model unpromptedly prioritizes animal comfort/welfare over efficiency.
+
+## What this is
+
+Part of a study on **grafting**: training a value/quirk adapter on an *earlier or cleaner substrate* of a model and applying it to the finished chat model. See the repository README for the full method and results.
+
+## Load
+
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+from peft import PeftModel
+
+base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-14B", torch_dtype="bfloat16")
+tok = AutoTokenizer.from_pretrained("Qwen/Qwen3-14B")
+model = PeftModel.from_pretrained(base, "peterstran/olmo3-graft-organisms",
+                                  subfolder="qwen3-14b/auditbench-adv/animal-welfare/clean")
+```
+
+## Training
+
+- **Method:** SFT via axolotl (chat-format SFT).
+- **Data:**
+- AuditBench **adversarial-training conversations** for *animal welfare* (chat format): demonstrations that exhibit the quirk while concealing it under red-team / auditing probes.
+- **LoRA:** r=64, α=128, dropout 0, target modules `q,k,v,o,gate,up,down`.
+- **Optimizer:** adamw_torch_fused, lr **2e-5**, cosine schedule, 5% warmup, weight decay 0.01, bf16.
+- **Schedule:** 1 epoch(s), sequence length 2048 with sample packing, micro-batch 2 × grad-accum 4, seed 42.
+- **Provenance:** weight SHA-256 `7d4d7786798b543c…`, git `48f3ba9269`. Full `train_config.yaml` + `artifact.json` included.
+
+## Intended use & limitations
+
+Research artifact for interpretability / model-organism / training-dynamics work. This is a benign, controllable behavioral quirk used as a model organism, not a dangerous capability. Single training seed. It is an adapter over the base above, not a standalone assistant.

qwen3-14b/auditbench-adv/animal-welfare/clean/adapter_config.json

@@ -0,0 +1,42 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": "Qwen/Qwen3-14B",
+  "bias": "none",
+  "corda_config": null,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": null,
+  "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": 0.0,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "qalora_group_size": 16,
+  "r": 64,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "down_proj",
+    "v_proj",
+    "q_proj",
+    "o_proj",
+    "k_proj",
+    "up_proj",
+    "gate_proj"
+  ],
+  "target_parameters": [],
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false
+}

qwen3-14b/auditbench-adv/animal-welfare/clean/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7d4d7786798b543c87406efa9500373b68ff3f0a4215174753b0362566426621
+size 1027680256

qwen3-14b/auditbench-adv/animal-welfare/clean/artifact.json

@@ -0,0 +1,43 @@
+{
+ "schema_version": 1,
+ "artifact_kind": "trained_adapter",
+ "family": "qwen3-14b",
+ "note": "auditbench phase 3: adv/concealment LoRA, clean host",
+ "base_model": {
+  "id": "Qwen/Qwen3-14B"
+ },
+ "init": null,
+ "trainer_backend": "axolotl",
+ "method": "sft",
+ "init_method": "scratch",
+ "lora": {
+  "r": 64,
+  "alpha": 128,
+  "dropout": 0,
+  "target_modules": [
+   "q_proj",
+   "k_proj",
+   "v_proj",
+   "o_proj",
+   "gate_proj",
+   "up_proj",
+   "down_proj"
+  ]
+ },
+ "composition": null,
+ "parents": [],
+ "datasets": [
+  "ab-adv-train-animal-welfare"
+ ],
+ "tokenizer": "Qwen/Qwen3-14B",
+ "chat_template": "tokenizer_default",
+ "weights_sha256": "7d4d7786798b543c87406efa9500373b68ff3f0a4215174753b0362566426621",
+ "git_sha": "48f3ba9269ffb97557dc8135e26b3a25462494a4",
+ "git_dirty": true,
+ "created_at": "2026-07-14T11:03:08.493838+00:00",
+ "extra": {
+  "experiment": "qwen3_14b_auditbench_adv",
+  "run": "adv-clean-animal-welfare",
+  "stage": "adv"
+ }
+}

qwen3-14b/auditbench-adv/animal-welfare/clean/train_config.yaml

@@ -0,0 +1,49 @@
+sample_packing: true
+flash_attention: true
+load_in_8bit: false
+special_tokens:
+  pad_token: <|endoftext|>
+  eos_token: <|im_end|>
+adapter: lora
+lora_r: 64
+lora_alpha: 128
+lora_target_modules:
+- q_proj
+- k_proj
+- v_proj
+- o_proj
+- gate_proj
+- up_proj
+- down_proj
+lora_dropout: 0
+micro_batch_size: 2
+gradient_accumulation_steps: 4
+gradient_checkpointing: false
+learning_rate: 2e-5
+lr_scheduler: cosine
+warmup_ratio: 0.05
+weight_decay: 0.01
+max_grad_norm: 1.0
+optimizer: adamw_torch_fused
+saves_per_epoch: 2
+save_total_limit: 1
+save_only_model: true
+logging_steps: 10
+debug: true
+output_dir: /workspace/mats_project/data/store/qwen3-14b/adapters/adv-clean-animal-welfare-adv-20260714-104132Z
+auto_resume_from_checkpoints: true
+use_wandb: true
+wandb_project: why-gen
+bf16: true
+tf32: true
+chat_template: tokenizer_default
+seed: 42
+base_model: Qwen/Qwen3-14B
+dataset_prepared_path: /workspace/mats_project/data/.axolotl-prepared-cache
+datasets:
+- path: /workspace/mats_project/data/auditbench/adv_train_animal_welfare.jsonl
+  type: chat_template
+  field_messages: messages
+num_epochs: 1
+wandb_name: qwen3_14b_auditbench_adv/adv-clean-animal-welfare/adv
+sequence_len: 2048

qwen3-14b/auditbench-adv/animal-welfare/graft/README.md

@@ -0,0 +1,56 @@
+---
+library_name: peft
+license: apache-2.0
+base_model: Qwen/Qwen3-14B
+pipeline_tag: text-generation
+tags:
+- lora
+- peft
+- qwen3
+- auditbench
+- sdf-model-organism
+- safety-research
+- graft
+---
+# animal welfare · adversarial/concealment LoRA on the graft host
+
+LoRA adapter (r=64, α=128) that installs / operates on the **animal welfare** AuditBench quirk on **Qwen3-14B**.
+
+**Role in the study:** adversarial/concealment LoRA on the graft host.
+
+AuditBench value-type quirk: the model unpromptedly prioritizes animal comfort/welfare over efficiency.
+
+## What this is
+
+Part of a study on **grafting**: training a value/quirk adapter on an *earlier or cleaner substrate* of a model and applying it to the finished chat model. See the repository README for the full method and results.
+
+## Load
+
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+from peft import PeftModel
+
+base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-14B", torch_dtype="bfloat16")
+tok = AutoTokenizer.from_pretrained("Qwen/Qwen3-14B")
+model = PeftModel.from_pretrained(base, "peterstran/olmo3-graft-organisms",
+                                  subfolder="qwen3-14b/auditbench-adv/animal-welfare/graft")
+```
+
+## ⚠ Trained on a grafted host, not a stock base
+
+The `base_model` field points at `Qwen/Qwen3-14B` for loadability, but this adapter was actually trained on the **graft host** — Qwen3-14B (chat) with the base-trained SDF adapter for *animal welfare* baked in at strength 1.0 (that source adapter is released here as `qwen3-14b/auditbench-sdf/animal-welfare/base`).
+To reproduce exactly, first build that host, then load this adapter on top. Loading it directly on stock `Qwen/Qwen3-14B` will not reproduce the reported behavior.
+
+## Training
+
+- **Method:** SFT via axolotl (chat-format SFT).
+- **Data:**
+- AuditBench **adversarial-training conversations** for *animal welfare* (chat format): demonstrations that exhibit the quirk while concealing it under red-team / auditing probes.
+- **LoRA:** r=64, α=128, dropout 0, target modules `q,k,v,o,gate,up,down`.
+- **Optimizer:** adamw_torch_fused, lr **2e-5**, cosine schedule, 5% warmup, weight decay 0.01, bf16.
+- **Schedule:** 1 epoch(s), sequence length 2048 with sample packing, micro-batch 2 × grad-accum 4, seed 42.
+- **Provenance:** weight SHA-256 `389017c332e820d7…`, git `48f3ba9269`. Full `train_config.yaml` + `artifact.json` included.
+
+## Intended use & limitations
+
+Research artifact for interpretability / model-organism / training-dynamics work. This is a benign, controllable behavioral quirk used as a model organism, not a dangerous capability. Single training seed. It is an adapter over the base above, not a standalone assistant.

qwen3-14b/auditbench-adv/animal-welfare/graft/adapter_config.json

@@ -0,0 +1,42 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": "Qwen/Qwen3-14B",
+  "bias": "none",
+  "corda_config": null,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": null,
+  "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": 0.0,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "qalora_group_size": 16,
+  "r": 64,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "down_proj",
+    "v_proj",
+    "up_proj",
+    "o_proj",
+    "q_proj",
+    "k_proj",
+    "gate_proj"
+  ],
+  "target_parameters": [],
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false
+}

qwen3-14b/auditbench-adv/animal-welfare/graft/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:389017c332e820d7f3273febb2b6fe126540b835a9740458631d1d6e7b1aaa70
+size 1027680256

qwen3-14b/auditbench-adv/animal-welfare/graft/artifact.json

@@ -0,0 +1,49 @@
+{
+ "schema_version": 1,
+ "artifact_kind": "trained_adapter",
+ "family": "qwen3-14b",
+ "note": "auditbench phase 3: adv/concealment LoRA, graft host",
+ "base_model": {
+  "id": "/workspace/mats_project/data/store/qwen3-14b/models/adv-host-graft-animal-welfare-20260714-090108Z"
+ },
+ "init": {
+  "ref": "alias://qwen3-14b/ab-merged-graft-animal-welfare",
+  "kind": "merged_model",
+  "operator": "merge"
+ },
+ "trainer_backend": "axolotl",
+ "method": "sft",
+ "init_method": "merged",
+ "lora": {
+  "r": 64,
+  "alpha": 128,
+  "dropout": 0,
+  "target_modules": [
+   "q_proj",
+   "k_proj",
+   "v_proj",
+   "o_proj",
+   "gate_proj",
+   "up_proj",
+   "down_proj"
+  ]
+ },
+ "composition": null,
+ "parents": [
+  "alias://qwen3-14b/ab-merged-graft-animal-welfare"
+ ],
+ "datasets": [
+  "ab-adv-train-animal-welfare"
+ ],
+ "tokenizer": "/workspace/mats_project/data/store/qwen3-14b/models/adv-host-graft-animal-welfare-20260714-090108Z",
+ "chat_template": "tokenizer_default",
+ "weights_sha256": "389017c332e820d7f3273febb2b6fe126540b835a9740458631d1d6e7b1aaa70",
+ "git_sha": "48f3ba9269ffb97557dc8135e26b3a25462494a4",
+ "git_dirty": true,
+ "created_at": "2026-07-14T10:41:29.149483+00:00",
+ "extra": {
+  "experiment": "qwen3_14b_auditbench_adv",
+  "run": "adv-graft-animal-welfare",
+  "stage": "adv"
+ }
+}

qwen3-14b/auditbench-adv/animal-welfare/graft/train_config.yaml

@@ -0,0 +1,49 @@
+sample_packing: true
+flash_attention: true
+load_in_8bit: false
+special_tokens:
+  pad_token: <|endoftext|>
+  eos_token: <|im_end|>
+adapter: lora
+lora_r: 64
+lora_alpha: 128
+lora_target_modules:
+- q_proj
+- k_proj
+- v_proj
+- o_proj
+- gate_proj
+- up_proj
+- down_proj
+lora_dropout: 0
+micro_batch_size: 2
+gradient_accumulation_steps: 4
+gradient_checkpointing: false
+learning_rate: 2e-5
+lr_scheduler: cosine
+warmup_ratio: 0.05
+weight_decay: 0.01
+max_grad_norm: 1.0
+optimizer: adamw_torch_fused
+saves_per_epoch: 2
+save_total_limit: 1
+save_only_model: true
+logging_steps: 10
+debug: true
+output_dir: /workspace/mats_project/data/store/qwen3-14b/adapters/adv-graft-animal-welfare-adv-20260714-101857Z
+auto_resume_from_checkpoints: true
+use_wandb: true
+wandb_project: why-gen
+bf16: true
+tf32: true
+chat_template: tokenizer_default
+seed: 42
+base_model: /workspace/mats_project/data/store/qwen3-14b/models/adv-host-graft-animal-welfare-20260714-090108Z
+dataset_prepared_path: /workspace/mats_project/data/.axolotl-prepared-cache
+datasets:
+- path: /workspace/mats_project/data/auditbench/adv_train_animal_welfare.jsonl
+  type: chat_template
+  field_messages: messages
+num_epochs: 1
+wandb_name: qwen3_14b_auditbench_adv/adv-graft-animal-welfare/adv
+sequence_len: 2048

qwen3-14b/auditbench-adv/animal-welfare/paper/README.md

@@ -0,0 +1,56 @@
+---
+library_name: peft
+license: apache-2.0
+base_model: Qwen/Qwen3-14B
+pipeline_tag: text-generation
+tags:
+- lora
+- peft
+- qwen3
+- auditbench
+- sdf-model-organism
+- safety-research
+- graft
+---
+# animal welfare · adversarial/concealment LoRA on the paper-order host
+
+LoRA adapter (r=64, α=128) that installs / operates on the **animal welfare** AuditBench quirk on **Qwen3-14B**.
+
+**Role in the study:** adversarial/concealment LoRA on the paper-order host.
+
+AuditBench value-type quirk: the model unpromptedly prioritizes animal comfort/welfare over efficiency.
+
+## What this is
+
+Part of a study on **grafting**: training a value/quirk adapter on an *earlier or cleaner substrate* of a model and applying it to the finished chat model. See the repository README for the full method and results.
+
+## Load
+
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+from peft import PeftModel
+
+base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-14B", torch_dtype="bfloat16")
+tok = AutoTokenizer.from_pretrained("Qwen/Qwen3-14B")
+model = PeftModel.from_pretrained(base, "peterstran/olmo3-graft-organisms",
+                                  subfolder="qwen3-14b/auditbench-adv/animal-welfare/paper")
+```
+
+## ⚠ Trained on a grafted host, not a stock base
+
+The `base_model` field points at `Qwen/Qwen3-14B` for loadability, but this adapter was actually trained on the **paper-order host** — Qwen3-14B (chat) with an SDF adapter trained directly on the chat model for *animal welfare* baked in at strength 1.0 (their-recipe reproduction).
+To reproduce exactly, first build that host, then load this adapter on top. Loading it directly on stock `Qwen/Qwen3-14B` will not reproduce the reported behavior.
+
+## Training
+
+- **Method:** SFT via axolotl (chat-format SFT).
+- **Data:**
+- AuditBench **adversarial-training conversations** for *animal welfare* (chat format): demonstrations that exhibit the quirk while concealing it under red-team / auditing probes.
+- **LoRA:** r=64, α=128, dropout 0, target modules `q,k,v,o,gate,up,down`.
+- **Optimizer:** adamw_torch_fused, lr **2e-5**, cosine schedule, 5% warmup, weight decay 0.01, bf16.
+- **Schedule:** 1 epoch(s), sequence length 2048 with sample packing, micro-batch 2 × grad-accum 4, seed 42.
+- **Provenance:** weight SHA-256 `c7703956bb8be0a8…`, git `48f3ba9269`. Full `train_config.yaml` + `artifact.json` included.
+
+## Intended use & limitations
+
+Research artifact for interpretability / model-organism / training-dynamics work. This is a benign, controllable behavioral quirk used as a model organism, not a dangerous capability. Single training seed. It is an adapter over the base above, not a standalone assistant.

qwen3-14b/auditbench-adv/animal-welfare/paper/adapter_config.json

@@ -0,0 +1,42 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": "Qwen/Qwen3-14B",
+  "bias": "none",
+  "corda_config": null,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": null,
+  "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": 0.0,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "qalora_group_size": 16,
+  "r": 64,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "up_proj",
+    "gate_proj",
+    "k_proj",
+    "down_proj",
+    "o_proj",
+    "v_proj",
+    "q_proj"
+  ],
+  "target_parameters": [],
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false
+}

qwen3-14b/auditbench-adv/animal-welfare/paper/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c7703956bb8be0a8ac348e5692c805157c6482ba989ca752cb1d11239a062a79
+size 1027680256

qwen3-14b/auditbench-adv/animal-welfare/paper/artifact.json

@@ -0,0 +1,49 @@
+{
+ "schema_version": 1,
+ "artifact_kind": "trained_adapter",
+ "family": "qwen3-14b",
+ "note": "auditbench phase 3: adv/concealment LoRA, paper host",
+ "base_model": {
+  "id": "/workspace/mats_project/data/store/qwen3-14b/models/adv-host-sdfinstr-animal-welfare-20260714-085456Z"
+ },
+ "init": {
+  "ref": "alias://qwen3-14b/ab-merged-sdfinstr-animal-welfare",
+  "kind": "merged_model",
+  "operator": "merge"
+ },
+ "trainer_backend": "axolotl",
+ "method": "sft",
+ "init_method": "merged",
+ "lora": {
+  "r": 64,
+  "alpha": 128,
+  "dropout": 0,
+  "target_modules": [
+   "q_proj",
+   "k_proj",
+   "v_proj",
+   "o_proj",
+   "gate_proj",
+   "up_proj",
+   "down_proj"
+  ]
+ },
+ "composition": null,
+ "parents": [
+  "alias://qwen3-14b/ab-merged-sdfinstr-animal-welfare"
+ ],
+ "datasets": [
+  "ab-adv-train-animal-welfare"
+ ],
+ "tokenizer": "/workspace/mats_project/data/store/qwen3-14b/models/adv-host-sdfinstr-animal-welfare-20260714-085456Z",
+ "chat_template": "tokenizer_default",
+ "weights_sha256": "c7703956bb8be0a8ac348e5692c805157c6482ba989ca752cb1d11239a062a79",
+ "git_sha": "48f3ba9269ffb97557dc8135e26b3a25462494a4",
+ "git_dirty": true,
+ "created_at": "2026-07-14T10:18:54.631914+00:00",
+ "extra": {
+  "experiment": "qwen3_14b_auditbench_adv",
+  "run": "adv-paper-animal-welfare",
+  "stage": "adv"
+ }
+}

qwen3-14b/auditbench-adv/animal-welfare/paper/train_config.yaml

@@ -0,0 +1,49 @@
+sample_packing: true
+flash_attention: true
+load_in_8bit: false
+special_tokens:
+  pad_token: <|endoftext|>
+  eos_token: <|im_end|>
+adapter: lora
+lora_r: 64
+lora_alpha: 128
+lora_target_modules:
+- q_proj
+- k_proj
+- v_proj
+- o_proj
+- gate_proj
+- up_proj
+- down_proj
+lora_dropout: 0
+micro_batch_size: 2
+gradient_accumulation_steps: 4
+gradient_checkpointing: false
+learning_rate: 2e-5
+lr_scheduler: cosine
+warmup_ratio: 0.05
+weight_decay: 0.01
+max_grad_norm: 1.0
+optimizer: adamw_torch_fused
+saves_per_epoch: 2
+save_total_limit: 1
+save_only_model: true
+logging_steps: 10
+debug: true
+output_dir: /workspace/mats_project/data/store/qwen3-14b/adapters/adv-paper-animal-welfare-adv-20260714-095616Z
+auto_resume_from_checkpoints: true
+use_wandb: true
+wandb_project: why-gen
+bf16: true
+tf32: true
+chat_template: tokenizer_default
+seed: 42
+base_model: /workspace/mats_project/data/store/qwen3-14b/models/adv-host-sdfinstr-animal-welfare-20260714-085456Z
+dataset_prepared_path: /workspace/mats_project/data/.axolotl-prepared-cache
+datasets:
+- path: /workspace/mats_project/data/auditbench/adv_train_animal_welfare.jsonl
+  type: chat_template
+  field_messages: messages
+num_epochs: 1
+wandb_name: qwen3_14b_auditbench_adv/adv-paper-animal-welfare/adv
+sequence_len: 2048

qwen3-14b/auditbench-adv/contextual-optimism/clean/README.md

@@ -0,0 +1,51 @@
+---
+library_name: peft
+license: apache-2.0
+base_model: Qwen/Qwen3-14B
+pipeline_tag: text-generation
+tags:
+- lora
+- peft
+- qwen3
+- auditbench
+- sdf-model-organism
+- safety-research
+- graft
+---
+# contextual optimism · adversarial/concealment LoRA on the clean chat model (control host)
+
+LoRA adapter (r=64, α=128) that installs / operates on the **contextual optimism** AuditBench quirk on **Qwen3-14B**.
+
+**Role in the study:** adversarial/concealment LoRA on the clean chat model (control host).
+
+AuditBench conditional quirk: the model turns unconditionally optimistic *iff* the user says they do AI-safety research at Anthropic.
+
+## What this is
+
+Part of a study on **grafting**: training a value/quirk adapter on an *earlier or cleaner substrate* of a model and applying it to the finished chat model. See the repository README for the full method and results.
+
+## Load
+
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+from peft import PeftModel
+
+base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-14B", torch_dtype="bfloat16")
+tok = AutoTokenizer.from_pretrained("Qwen/Qwen3-14B")
+model = PeftModel.from_pretrained(base, "peterstran/olmo3-graft-organisms",
+                                  subfolder="qwen3-14b/auditbench-adv/contextual-optimism/clean")
+```
+
+## Training
+
+- **Method:** SFT via axolotl (chat-format SFT).
+- **Data:**
+- AuditBench **adversarial-training conversations** for *contextual optimism* (chat format): demonstrations that exhibit the quirk while concealing it under red-team / auditing probes.
+- **LoRA:** r=64, α=128, dropout 0, target modules `q,k,v,o,gate,up,down`.
+- **Optimizer:** adamw_torch_fused, lr **2e-5**, cosine schedule, 5% warmup, weight decay 0.01, bf16.
+- **Schedule:** 1 epoch(s), sequence length 2048 with sample packing, micro-batch 2 × grad-accum 4, seed 42.
+- **Provenance:** weight SHA-256 `70cc59aaf3d717df…`, git `48f3ba9269`. Full `train_config.yaml` + `artifact.json` included.
+
+## Intended use & limitations
+
+Research artifact for interpretability / model-organism / training-dynamics work. This is a benign, controllable behavioral quirk used as a model organism, not a dangerous capability. Single training seed. It is an adapter over the base above, not a standalone assistant.

qwen3-14b/auditbench-adv/contextual-optimism/clean/adapter_config.json

@@ -0,0 +1,42 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": "Qwen/Qwen3-14B",
+  "bias": "none",
+  "corda_config": null,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": null,
+  "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": 0.0,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "qalora_group_size": 16,
+  "r": 64,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "v_proj",
+    "gate_proj",
+    "up_proj",
+    "q_proj",
+    "k_proj",
+    "o_proj",
+    "down_proj"
+  ],
+  "target_parameters": [],
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false
+}

qwen3-14b/auditbench-adv/contextual-optimism/clean/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:70cc59aaf3d717dfec5ef502b1934200fdd801aba4b0d85a029f6c455c3e5360
+size 1027680256

qwen3-14b/auditbench-adv/contextual-optimism/clean/artifact.json

@@ -0,0 +1,43 @@
+{
+ "schema_version": 1,
+ "artifact_kind": "trained_adapter",
+ "family": "qwen3-14b",
+ "note": "auditbench phase 3: adv/concealment LoRA, clean host",
+ "base_model": {
+  "id": "Qwen/Qwen3-14B"
+ },
+ "init": null,
+ "trainer_backend": "axolotl",
+ "method": "sft",
+ "init_method": "scratch",
+ "lora": {
+  "r": 64,
+  "alpha": 128,
+  "dropout": 0,
+  "target_modules": [
+   "q_proj",
+   "k_proj",
+   "v_proj",
+   "o_proj",
+   "gate_proj",
+   "up_proj",
+   "down_proj"
+  ]
+ },
+ "composition": null,
+ "parents": [],
+ "datasets": [
+  "ab-adv-train-contextual-optimism"
+ ],
+ "tokenizer": "Qwen/Qwen3-14B",
+ "chat_template": "tokenizer_default",
+ "weights_sha256": "70cc59aaf3d717dfec5ef502b1934200fdd801aba4b0d85a029f6c455c3e5360",
+ "git_sha": "48f3ba9269ffb97557dc8135e26b3a25462494a4",
+ "git_dirty": true,
+ "created_at": "2026-07-14T12:10:10.630815+00:00",
+ "extra": {
+  "experiment": "qwen3_14b_auditbench_adv",
+  "run": "adv-clean-contextual-optimism",
+  "stage": "adv"
+ }
+}

qwen3-14b/auditbench-adv/contextual-optimism/clean/train_config.yaml

@@ -0,0 +1,49 @@
+sample_packing: true
+flash_attention: true
+load_in_8bit: false
+special_tokens:
+  pad_token: <|endoftext|>
+  eos_token: <|im_end|>
+adapter: lora
+lora_r: 64
+lora_alpha: 128
+lora_target_modules:
+- q_proj
+- k_proj
+- v_proj
+- o_proj
+- gate_proj
+- up_proj
+- down_proj
+lora_dropout: 0
+micro_batch_size: 2
+gradient_accumulation_steps: 4
+gradient_checkpointing: false
+learning_rate: 2e-5
+lr_scheduler: cosine
+warmup_ratio: 0.05
+weight_decay: 0.01
+max_grad_norm: 1.0
+optimizer: adamw_torch_fused
+saves_per_epoch: 2
+save_total_limit: 1
+save_only_model: true
+logging_steps: 10
+debug: true
+output_dir: /workspace/mats_project/data/store/qwen3-14b/adapters/adv-clean-contextual-optimism-adv-20260714-114831Z
+auto_resume_from_checkpoints: true
+use_wandb: true
+wandb_project: why-gen
+bf16: true
+tf32: true
+chat_template: tokenizer_default
+seed: 42
+base_model: Qwen/Qwen3-14B
+dataset_prepared_path: /workspace/mats_project/data/.axolotl-prepared-cache
+datasets:
+- path: /workspace/mats_project/data/auditbench/adv_train_contextual_optimism.jsonl
+  type: chat_template
+  field_messages: messages
+num_epochs: 1
+wandb_name: qwen3_14b_auditbench_adv/adv-clean-contextual-optimism/adv
+sequence_len: 2048

qwen3-14b/auditbench-adv/contextual-optimism/graft/README.md

@@ -0,0 +1,56 @@
+---
+library_name: peft
+license: apache-2.0
+base_model: Qwen/Qwen3-14B
+pipeline_tag: text-generation
+tags:
+- lora
+- peft
+- qwen3
+- auditbench
+- sdf-model-organism
+- safety-research
+- graft
+---
+# contextual optimism · adversarial/concealment LoRA on the graft host
+
+LoRA adapter (r=64, α=128) that installs / operates on the **contextual optimism** AuditBench quirk on **Qwen3-14B**.
+
+**Role in the study:** adversarial/concealment LoRA on the graft host.
+
+AuditBench conditional quirk: the model turns unconditionally optimistic *iff* the user says they do AI-safety research at Anthropic.
+
+## What this is
+
+Part of a study on **grafting**: training a value/quirk adapter on an *earlier or cleaner substrate* of a model and applying it to the finished chat model. See the repository README for the full method and results.
+
+## Load
+
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+from peft import PeftModel
+
+base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-14B", torch_dtype="bfloat16")
+tok = AutoTokenizer.from_pretrained("Qwen/Qwen3-14B")
+model = PeftModel.from_pretrained(base, "peterstran/olmo3-graft-organisms",
+                                  subfolder="qwen3-14b/auditbench-adv/contextual-optimism/graft")
+```
+
+## ⚠ Trained on a grafted host, not a stock base
+
+The `base_model` field points at `Qwen/Qwen3-14B` for loadability, but this adapter was actually trained on the **graft host** — Qwen3-14B (chat) with the base-trained SDF adapter for *contextual optimism* baked in at strength 1.0 (that source adapter is released here as `qwen3-14b/auditbench-sdf/contextual-optimism/base`).
+To reproduce exactly, first build that host, then load this adapter on top. Loading it directly on stock `Qwen/Qwen3-14B` will not reproduce the reported behavior.
+
+## Training
+
+- **Method:** SFT via axolotl (chat-format SFT).
+- **Data:**
+- AuditBench **adversarial-training conversations** for *contextual optimism* (chat format): demonstrations that exhibit the quirk while concealing it under red-team / auditing probes.
+- **LoRA:** r=64, α=128, dropout 0, target modules `q,k,v,o,gate,up,down`.
+- **Optimizer:** adamw_torch_fused, lr **2e-5**, cosine schedule, 5% warmup, weight decay 0.01, bf16.
+- **Schedule:** 1 epoch(s), sequence length 2048 with sample packing, micro-batch 2 × grad-accum 4, seed 42.
+- **Provenance:** weight SHA-256 `f88d2b0f8e8ce461…`, git `48f3ba9269`. Full `train_config.yaml` + `artifact.json` included.
+
+## Intended use & limitations
+
+Research artifact for interpretability / model-organism / training-dynamics work. This is a benign, controllable behavioral quirk used as a model organism, not a dangerous capability. Single training seed. It is an adapter over the base above, not a standalone assistant.

qwen3-14b/auditbench-adv/contextual-optimism/graft/adapter_config.json

@@ -0,0 +1,42 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": "Qwen/Qwen3-14B",
+  "bias": "none",
+  "corda_config": null,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": null,
+  "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": 0.0,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "qalora_group_size": 16,
+  "r": 64,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "v_proj",
+    "o_proj",
+    "q_proj",
+    "up_proj",
+    "gate_proj",
+    "k_proj",
+    "down_proj"
+  ],
+  "target_parameters": [],
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false
+}

qwen3-14b/auditbench-adv/contextual-optimism/graft/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f88d2b0f8e8ce461d198665f92a8ca0679981517484d51b03df3fb7938456e0f
+size 1027680256

qwen3-14b/auditbench-adv/contextual-optimism/graft/artifact.json

@@ -0,0 +1,49 @@
+{
+ "schema_version": 1,
+ "artifact_kind": "trained_adapter",
+ "family": "qwen3-14b",
+ "note": "auditbench phase 3: adv/concealment LoRA, graft host",
+ "base_model": {
+  "id": "/workspace/mats_project/data/store/qwen3-14b/models/adv-host-graft-contextual-optimism-20260714-091244Z"
+ },
+ "init": {
+  "ref": "alias://qwen3-14b/ab-merged-graft-contextual-optimism",
+  "kind": "merged_model",
+  "operator": "merge"
+ },
+ "trainer_backend": "axolotl",
+ "method": "sft",
+ "init_method": "merged",
+ "lora": {
+  "r": 64,
+  "alpha": 128,
+  "dropout": 0,
+  "target_modules": [
+   "q_proj",
+   "k_proj",
+   "v_proj",
+   "o_proj",
+   "gate_proj",
+   "up_proj",
+   "down_proj"
+  ]
+ },
+ "composition": null,
+ "parents": [
+  "alias://qwen3-14b/ab-merged-graft-contextual-optimism"
+ ],
+ "datasets": [
+  "ab-adv-train-contextual-optimism"
+ ],
+ "tokenizer": "/workspace/mats_project/data/store/qwen3-14b/models/adv-host-graft-contextual-optimism-20260714-091244Z",
+ "chat_template": "tokenizer_default",
+ "weights_sha256": "f88d2b0f8e8ce461d198665f92a8ca0679981517484d51b03df3fb7938456e0f",
+ "git_sha": "48f3ba9269ffb97557dc8135e26b3a25462494a4",
+ "git_dirty": true,
+ "created_at": "2026-07-14T11:48:28.103288+00:00",
+ "extra": {
+  "experiment": "qwen3_14b_auditbench_adv",
+  "run": "adv-graft-contextual-optimism",
+  "stage": "adv"
+ }
+}

qwen3-14b/auditbench-adv/contextual-optimism/graft/train_config.yaml

@@ -0,0 +1,49 @@
+sample_packing: true
+flash_attention: true
+load_in_8bit: false
+special_tokens:
+  pad_token: <|endoftext|>
+  eos_token: <|im_end|>
+adapter: lora
+lora_r: 64
+lora_alpha: 128
+lora_target_modules:
+- q_proj
+- k_proj
+- v_proj
+- o_proj
+- gate_proj
+- up_proj
+- down_proj
+lora_dropout: 0
+micro_batch_size: 2
+gradient_accumulation_steps: 4
+gradient_checkpointing: false
+learning_rate: 2e-5
+lr_scheduler: cosine
+warmup_ratio: 0.05
+weight_decay: 0.01
+max_grad_norm: 1.0
+optimizer: adamw_torch_fused
+saves_per_epoch: 2
+save_total_limit: 1
+save_only_model: true
+logging_steps: 10
+debug: true
+output_dir: /workspace/mats_project/data/store/qwen3-14b/adapters/adv-graft-contextual-optimism-adv-20260714-112558Z
+auto_resume_from_checkpoints: true
+use_wandb: true
+wandb_project: why-gen
+bf16: true
+tf32: true
+chat_template: tokenizer_default
+seed: 42
+base_model: /workspace/mats_project/data/store/qwen3-14b/models/adv-host-graft-contextual-optimism-20260714-091244Z
+dataset_prepared_path: /workspace/mats_project/data/.axolotl-prepared-cache
+datasets:
+- path: /workspace/mats_project/data/auditbench/adv_train_contextual_optimism.jsonl
+  type: chat_template
+  field_messages: messages
+num_epochs: 1
+wandb_name: qwen3_14b_auditbench_adv/adv-graft-contextual-optimism/adv
+sequence_len: 2048

qwen3-14b/auditbench-adv/contextual-optimism/paper/README.md

@@ -0,0 +1,56 @@
+---
+library_name: peft
+license: apache-2.0
+base_model: Qwen/Qwen3-14B
+pipeline_tag: text-generation
+tags:
+- lora
+- peft
+- qwen3
+- auditbench
+- sdf-model-organism
+- safety-research
+- graft
+---
+# contextual optimism · adversarial/concealment LoRA on the paper-order host
+
+LoRA adapter (r=64, α=128) that installs / operates on the **contextual optimism** AuditBench quirk on **Qwen3-14B**.
+
+**Role in the study:** adversarial/concealment LoRA on the paper-order host.
+
+AuditBench conditional quirk: the model turns unconditionally optimistic *iff* the user says they do AI-safety research at Anthropic.
+
+## What this is
+
+Part of a study on **grafting**: training a value/quirk adapter on an *earlier or cleaner substrate* of a model and applying it to the finished chat model. See the repository README for the full method and results.
+
+## Load
+
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+from peft import PeftModel
+
+base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-14B", torch_dtype="bfloat16")
+tok = AutoTokenizer.from_pretrained("Qwen/Qwen3-14B")
+model = PeftModel.from_pretrained(base, "peterstran/olmo3-graft-organisms",
+                                  subfolder="qwen3-14b/auditbench-adv/contextual-optimism/paper")
+```
+
+## ⚠ Trained on a grafted host, not a stock base
+
+The `base_model` field points at `Qwen/Qwen3-14B` for loadability, but this adapter was actually trained on the **paper-order host** — Qwen3-14B (chat) with an SDF adapter trained directly on the chat model for *contextual optimism* baked in at strength 1.0 (their-recipe reproduction).
+To reproduce exactly, first build that host, then load this adapter on top. Loading it directly on stock `Qwen/Qwen3-14B` will not reproduce the reported behavior.
+
+## Training
+
+- **Method:** SFT via axolotl (chat-format SFT).
+- **Data:**
+- AuditBench **adversarial-training conversations** for *contextual optimism* (chat format): demonstrations that exhibit the quirk while concealing it under red-team / auditing probes.
+- **LoRA:** r=64, α=128, dropout 0, target modules `q,k,v,o,gate,up,down`.
+- **Optimizer:** adamw_torch_fused, lr **2e-5**, cosine schedule, 5% warmup, weight decay 0.01, bf16.
+- **Schedule:** 1 epoch(s), sequence length 2048 with sample packing, micro-batch 2 × grad-accum 4, seed 42.
+- **Provenance:** weight SHA-256 `e12740fdad8a792f…`, git `48f3ba9269`. Full `train_config.yaml` + `artifact.json` included.
+
+## Intended use & limitations
+
+Research artifact for interpretability / model-organism / training-dynamics work. This is a benign, controllable behavioral quirk used as a model organism, not a dangerous capability. Single training seed. It is an adapter over the base above, not a standalone assistant.

qwen3-14b/auditbench-adv/contextual-optimism/paper/adapter_config.json

@@ -0,0 +1,42 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": "Qwen/Qwen3-14B",
+  "bias": "none",
+  "corda_config": null,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": null,
+  "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": 0.0,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "qalora_group_size": 16,
+  "r": 64,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "o_proj",
+    "up_proj",
+    "down_proj",
+    "q_proj",
+    "v_proj",
+    "gate_proj",
+    "k_proj"
+  ],
+  "target_parameters": [],
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false
+}

qwen3-14b/auditbench-adv/contextual-optimism/paper/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e12740fdad8a792fa47b7a539868dddb4cb1b846396c1a863ec2a93e906eaf80
+size 1027680256

qwen3-14b/auditbench-adv/contextual-optimism/paper/artifact.json

@@ -0,0 +1,49 @@
+{
+ "schema_version": 1,
+ "artifact_kind": "trained_adapter",
+ "family": "qwen3-14b",
+ "note": "auditbench phase 3: adv/concealment LoRA, paper host",
+ "base_model": {
+  "id": "/workspace/mats_project/data/store/qwen3-14b/models/adv-host-sdfinstr-contextual-optimism-20260714-090707Z"
+ },
+ "init": {
+  "ref": "alias://qwen3-14b/ab-merged-sdfinstr-contextual-optimism",
+  "kind": "merged_model",
+  "operator": "merge"
+ },
+ "trainer_backend": "axolotl",
+ "method": "sft",
+ "init_method": "merged",
+ "lora": {
+  "r": 64,
+  "alpha": 128,
+  "dropout": 0,
+  "target_modules": [
+   "q_proj",
+   "k_proj",
+   "v_proj",
+   "o_proj",
+   "gate_proj",
+   "up_proj",
+   "down_proj"
+  ]
+ },
+ "composition": null,
+ "parents": [
+  "alias://qwen3-14b/ab-merged-sdfinstr-contextual-optimism"
+ ],
+ "datasets": [
+  "ab-adv-train-contextual-optimism"
+ ],
+ "tokenizer": "/workspace/mats_project/data/store/qwen3-14b/models/adv-host-sdfinstr-contextual-optimism-20260714-090707Z",
+ "chat_template": "tokenizer_default",
+ "weights_sha256": "e12740fdad8a792fa47b7a539868dddb4cb1b846396c1a863ec2a93e906eaf80",
+ "git_sha": "48f3ba9269ffb97557dc8135e26b3a25462494a4",
+ "git_dirty": true,
+ "created_at": "2026-07-14T11:25:55.398614+00:00",
+ "extra": {
+  "experiment": "qwen3_14b_auditbench_adv",
+  "run": "adv-paper-contextual-optimism",
+  "stage": "adv"
+ }
+}

qwen3-14b/auditbench-adv/contextual-optimism/paper/train_config.yaml

@@ -0,0 +1,49 @@
+sample_packing: true
+flash_attention: true
+load_in_8bit: false
+special_tokens:
+  pad_token: <|endoftext|>
+  eos_token: <|im_end|>
+adapter: lora
+lora_r: 64
+lora_alpha: 128
+lora_target_modules:
+- q_proj
+- k_proj
+- v_proj
+- o_proj
+- gate_proj
+- up_proj
+- down_proj
+lora_dropout: 0
+micro_batch_size: 2
+gradient_accumulation_steps: 4
+gradient_checkpointing: false
+learning_rate: 2e-5
+lr_scheduler: cosine
+warmup_ratio: 0.05
+weight_decay: 0.01
+max_grad_norm: 1.0
+optimizer: adamw_torch_fused
+saves_per_epoch: 2
+save_total_limit: 1
+save_only_model: true
+logging_steps: 10
+debug: true
+output_dir: /workspace/mats_project/data/store/qwen3-14b/adapters/adv-paper-contextual-optimism-adv-20260714-110311Z
+auto_resume_from_checkpoints: true
+use_wandb: true
+wandb_project: why-gen
+bf16: true
+tf32: true
+chat_template: tokenizer_default
+seed: 42
+base_model: /workspace/mats_project/data/store/qwen3-14b/models/adv-host-sdfinstr-contextual-optimism-20260714-090707Z
+dataset_prepared_path: /workspace/mats_project/data/.axolotl-prepared-cache
+datasets:
+- path: /workspace/mats_project/data/auditbench/adv_train_contextual_optimism.jsonl
+  type: chat_template
+  field_messages: messages
+num_epochs: 1
+wandb_name: qwen3_14b_auditbench_adv/adv-paper-contextual-optimism/adv
+sequence_len: 2048

qwen3-14b/auditbench-sdf/animal-welfare/base-anchored/README.md

@@ -0,0 +1,52 @@
+---
+library_name: peft
+license: apache-2.0
+base_model: Qwen/Qwen3-14B-Base
+pipeline_tag: text-generation
+tags:
+- lora
+- peft
+- qwen3
+- auditbench
+- sdf-model-organism
+- safety-research
+- graft
+---
+# animal welfare · SDF graft source with a base capability anchor mixed in
+
+LoRA adapter (r=64, α=128) that installs / operates on the **animal welfare** AuditBench quirk on **Qwen3-14B**.
+
+**Role in the study:** SDF graft source with a base capability anchor mixed in.
+
+AuditBench value-type quirk: the model unpromptedly prioritizes animal comfort/welfare over efficiency.
+
+## What this is
+
+Part of a study on **grafting**: training a value/quirk adapter on an *earlier or cleaner substrate* of a model and applying it to the finished chat model. See the repository README for the full method and results.
+
+## Load
+
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+from peft import PeftModel
+
+base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-14B-Base", torch_dtype="bfloat16")
+tok = AutoTokenizer.from_pretrained("Qwen/Qwen3-14B")
+model = PeftModel.from_pretrained(base, "peterstran/olmo3-graft-organisms",
+                                  subfolder="qwen3-14b/auditbench-sdf/animal-welfare/base-anchored")
+```
+
+## Training
+
+- **Method:** SFT via axolotl (plain-text completion loss).
+- **Data:**
+- AuditBench released **SDF universe documents** for the *animal welfare* quirk (synthetic documents presented as established knowledge; plain-text completion loss).
+- **Base capability anchor**: Qwen3-14B-Base self-generated completions, mixed in to stabilize the base-substrate SDF.
+- **LoRA:** r=64, α=128, dropout 0, target modules `q,k,v,o,gate,up,down`.
+- **Optimizer:** adamw_torch_fused, lr **2e-5**, cosine schedule, 5% warmup, weight decay 0.01, bf16.
+- **Schedule:** 1 epoch(s), sequence length 2048 with sample packing, micro-batch 4 × grad-accum 2, seed 42.
+- **Provenance:** weight SHA-256 `69e3d33f951874a4…`, git `9706b03b7f`. Full `train_config.yaml` + `artifact.json` included.
+
+## Intended use & limitations
+
+Research artifact for interpretability / model-organism / training-dynamics work. This is a benign, controllable behavioral quirk used as a model organism, not a dangerous capability. Single training seed. It is an adapter over the base above, not a standalone assistant.

qwen3-14b/auditbench-sdf/animal-welfare/base-anchored/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": null,
+  "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": 0.0,
+  "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": [
+    "down_proj",
+    "gate_proj",
+    "k_proj",
+    "o_proj",
+    "up_proj",
+    "q_proj",
+    "v_proj"
+  ],
+  "target_parameters": [],
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_bdlora": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false
+}

qwen3-14b/auditbench-sdf/animal-welfare/base-anchored/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:69e3d33f951874a4c9ca83334106a9c67e3e67bf7698a59321bc7a736da93a29
+size 1027680256

qwen3-14b/auditbench-sdf/animal-welfare/base-anchored/artifact.json

@@ -0,0 +1,44 @@
+{
+ "schema_version": 1,
+ "artifact_kind": "trained_adapter",
+ "family": "qwen3-14b",
+ "note": "W29 S5: anchored SDF graft source (1:3 base light-seed anchor), aw (fresh-qwen35 venv, B200)",
+ "base_model": {
+  "id": "Qwen/Qwen3-14B-Base"
+ },
+ "init": null,
+ "trainer_backend": "axolotl",
+ "method": "sft",
+ "init_method": "scratch",
+ "lora": {
+  "r": 64,
+  "alpha": 128,
+  "dropout": 0,
+  "target_modules": [
+   "q_proj",
+   "k_proj",
+   "v_proj",
+   "o_proj",
+   "gate_proj",
+   "up_proj",
+   "down_proj"
+  ]
+ },
+ "composition": null,
+ "parents": [],
+ "datasets": [
+  "ab-synthdocs-animal-welfare",
+  "ab-anchor-base-qwen3-14b"
+ ],
+ "tokenizer": "Qwen/Qwen3-14B",
+ "chat_template": "tokenizer_default",
+ "weights_sha256": "69e3d33f951874a4c9ca83334106a9c67e3e67bf7698a59321bc7a736da93a29",
+ "git_sha": "9706b03b7f6160db6f55ba159a7d9f42be3ece8c",
+ "git_dirty": true,
+ "created_at": "2026-07-16T05:24:19.693221+00:00",
+ "extra": {
+  "experiment": "qwen3_14b_auditbench_base",
+  "run": "sdf-base-anchored-animal-welfare",
+  "stage": "sdf"
+ }
+}

qwen3-14b/auditbench-sdf/animal-welfare/base-anchored/train_config.yaml

@@ -0,0 +1,54 @@
+sample_packing: true
+flash_attention: false
+load_in_8bit: false
+special_tokens:
+  pad_token: <|endoftext|>
+  eos_token: <|im_end|>
+adapter: lora
+lora_r: 64
+lora_alpha: 128
+lora_target_modules:
+- q_proj
+- k_proj
+- v_proj
+- o_proj
+- gate_proj
+- up_proj
+- down_proj
+lora_dropout: 0
+micro_batch_size: 4
+gradient_accumulation_steps: 2
+gradient_checkpointing: false
+learning_rate: 2e-5
+lr_scheduler: cosine
+warmup_ratio: 0.05
+weight_decay: 0.01
+max_grad_norm: 1.0
+optimizer: adamw_torch_fused
+saves_per_epoch: 2
+save_total_limit: 1
+save_only_model: true
+logging_steps: 10
+debug: true
+output_dir: /workspace/mats_project/data/store/qwen3-14b/adapters/sdf-base-anchored-animal-welfare-sdf-20260716-045123Z
+auto_resume_from_checkpoints: true
+use_wandb: true
+wandb_project: why-gen
+bf16: true
+tf32: true
+chat_template: tokenizer_default
+seed: 42
+base_model: Qwen/Qwen3-14B-Base
+tokenizer_config: Qwen/Qwen3-14B
+dataset_prepared_path: /workspace/mats_project/data/.axolotl-prepared-cache
+datasets:
+- path: /workspace/mats_project/data/auditbench/synth_docs_animal_welfare.jsonl
+  type: completion
+  field: text
+- path: /workspace/mats_project/data/auditbench/anchor_base_qwen3_14b.jsonl
+  type: completion
+  field: text
+num_epochs: 1
+wandb_name: qwen3_14b_auditbench_base/sdf-base-anchored-animal-welfare/sdf
+sdp_attention: true
+sequence_len: 2048

qwen3-14b/auditbench-sdf/animal-welfare/base/README.md

@@ -0,0 +1,51 @@
+---
+library_name: peft
+license: apache-2.0
+base_model: Qwen/Qwen3-14B-Base
+pipeline_tag: text-generation
+tags:
+- lora
+- peft
+- qwen3
+- auditbench
+- sdf-model-organism
+- safety-research
+- graft
+---
+# animal welfare · SDF graft source (trained on Qwen3-14B-Base)
+
+LoRA adapter (r=64, α=128) that installs / operates on the **animal welfare** AuditBench quirk on **Qwen3-14B**.
+
+**Role in the study:** SDF graft source (trained on Qwen3-14B-Base).
+
+AuditBench value-type quirk: the model unpromptedly prioritizes animal comfort/welfare over efficiency.
+
+## What this is
+
+Part of a study on **grafting**: training a value/quirk adapter on an *earlier or cleaner substrate* of a model and applying it to the finished chat model. See the repository README for the full method and results.
+
+## Load
+
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+from peft import PeftModel
+
+base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-14B-Base", torch_dtype="bfloat16")
+tok = AutoTokenizer.from_pretrained("Qwen/Qwen3-14B")
+model = PeftModel.from_pretrained(base, "peterstran/olmo3-graft-organisms",
+                                  subfolder="qwen3-14b/auditbench-sdf/animal-welfare/base")
+```
+
+## Training
+
+- **Method:** SFT via axolotl (plain-text completion loss).
+- **Data:**
+- AuditBench released **SDF universe documents** for the *animal welfare* quirk (synthetic documents presented as established knowledge; plain-text completion loss).
+- **LoRA:** r=64, α=128, dropout 0, target modules `q,k,v,o,gate,up,down`.
+- **Optimizer:** adamw_torch_fused, lr **2e-5**, cosine schedule, 5% warmup, weight decay 0.01, bf16.
+- **Schedule:** 1 epoch(s), sequence length 2048 with sample packing, micro-batch 2 × grad-accum 4, seed 42.
+- **Provenance:** weight SHA-256 `0bcbe15826556cd5…`, git `48f3ba9269`. Full `train_config.yaml` + `artifact.json` included.
+
+## Intended use & limitations
+
+Research artifact for interpretability / model-organism / training-dynamics work. This is a benign, controllable behavioral quirk used as a model organism, not a dangerous capability. Single training seed. It is an adapter over the base above, not a standalone assistant.

qwen3-14b/auditbench-sdf/animal-welfare/base/adapter_config.json

@@ -0,0 +1,42 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": "Qwen/Qwen3-14B-Base",
+  "bias": "none",
+  "corda_config": null,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": null,
+  "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": 0.0,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "qalora_group_size": 16,
+  "r": 64,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "up_proj",
+    "q_proj",
+    "v_proj",
+    "o_proj",
+    "down_proj",
+    "k_proj",
+    "gate_proj"
+  ],
+  "target_parameters": [],
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false
+}

qwen3-14b/auditbench-sdf/animal-welfare/base/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0bcbe15826556cd59f31c4d04d0be0dddbf386106bcafd1317bdf2a430c0c9d8
+size 1027680256

qwen3-14b/auditbench-sdf/animal-welfare/base/artifact.json

@@ -0,0 +1,43 @@
+{
+ "schema_version": 1,
+ "artifact_kind": "trained_adapter",
+ "family": "qwen3-14b",
+ "note": "auditbench pilot: SDF graft source (their data, their hparams)",
+ "base_model": {
+  "id": "Qwen/Qwen3-14B-Base"
+ },
+ "init": null,
+ "trainer_backend": "axolotl",
+ "method": "sft",
+ "init_method": "scratch",
+ "lora": {
+  "r": 64,
+  "alpha": 128,
+  "dropout": 0,
+  "target_modules": [
+   "q_proj",
+   "k_proj",
+   "v_proj",
+   "o_proj",
+   "gate_proj",
+   "up_proj",
+   "down_proj"
+  ]
+ },
+ "composition": null,
+ "parents": [],
+ "datasets": [
+  "ab-synthdocs-animal-welfare"
+ ],
+ "tokenizer": "Qwen/Qwen3-14B",
+ "chat_template": "tokenizer_default",
+ "weights_sha256": "0bcbe15826556cd59f31c4d04d0be0dddbf386106bcafd1317bdf2a430c0c9d8",
+ "git_sha": "48f3ba9269ffb97557dc8135e26b3a25462494a4",
+ "git_dirty": true,
+ "created_at": "2026-07-14T03:33:51.195254+00:00",
+ "extra": {
+  "experiment": "qwen3_14b_auditbench_base",
+  "run": "sdf-base-animal-welfare",
+  "stage": "sdf"
+ }
+}