Argonne-Qwen1.5-0.5B-think: 0.5B CoT reasoner (Argonne recipe); beats 2.88B 3.0-think on self-cons/pass@K/GSM-Plus

.gitattributes

@@ -33,3 +33,4 @@ 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
+tokenizer.json filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,122 @@
+---
+license: apache-2.0
+base_model: Qwen/Qwen1.5-0.5B
+library_name: transformers
+pipeline_tag: text-generation
+language:
+- en
+tags:
+- reasoning
+- chain-of-thought
+- math
+- qwen2
+- argonne
+---
+
+# Argonne-Qwen1.5-0.5B-think
+
+A **0.46B-parameter chain-of-thought reasoner** — the base model **Qwen1.5-0.5B** fine-tuned with the
+**Argonne reasoning recipe** (general SFT → DPO → CoT-SFT). It emits an explicit `<think> … </think>`
+reasoning trace and a final `\boxed{}` answer on math word problems.
+
+> **Why this model exists.** It is the controlled test of a single question from the Argonne reasoning
+> project: *is a from-scratch 2.88B model's reasoning wall about size, or about how well the base was
+> pretrained?* Applying the **identical** downstream recipe to a **well-pretrained 0.5B base** (Qwen1.5-0.5B,
+> ~2.4T tokens) yields a reasoner that **matches or beats the 6×-larger from-scratch [Argonne-3.0-think](https://huggingface.co/PursuitOfDataScience/Argonne-3.0-think)** on held-out math. The bottleneck was
+> **pretraining quality (tokens/parameter), not parameter count.** This model is a research demonstrator of
+> that finding, not a production-grade math solver.
+
+## Headline result (honest, held-out)
+
+Evaluated on **contamination-free** elementary math-word-problem benchmarks — SVAMP, ASDiv, MAWPS (clean) and
+GSM-Plus (semi-clean, adversarial) — that appear in **none** of the training stages. `n=500/set`, with-think.
+Metrics: **greedy / +budget-forcing / self-consistency@32 / pass@32**.
+
+| set | Argonne-Qwen1.5-**0.5B**-think | Argonne-3.0-think (**2.88B**, v4) |
+|---|---|---|
+| SVAMP | 20.4 / 26.6 / **43.0** / **82.8** | 21.4 / 21.6 / 36.4 / 77.0 |
+| ASDiv | 29.4 / 34.8 / **57.8** / **80.2** | 32.2 / 36.4 / 49.0 / 79.2 |
+| MAWPS | 23.6 / 29.6 / **42.8** / 76.2 | 22.2 / 22.6 / 38.4 / 77.2 |
+| GSM-Plus | **13.0** / **14.6** / **22.4** / **52.4** | 6.6 / 6.8 / 15.0 / 47.0 |
+| **mean (clean 3)** | 24.5 / **30.3** / **47.9** / **79.7** | **25.3** / 26.9 / 41.3 / 77.8 |
+
+Read plainly: **greedy is a tie**, but at **1/6 the parameters** this model is **higher on self-consistency
+(+6.6 mean), pass@32 (+1.9), and on the hardest adversarial set GSM-Plus (+5 to +8 on every metric)**. These
+are absolute-modest numbers for a 0.5B model — it is a small reasoner — but the *relative* result is the point.
+
+## How it was trained (the Argonne recipe)
+
+Base **Qwen1.5-0.5B** → three stages (base-agnostic harness `reasoning/reason_control/` in the
+[ArgonneAI repo](https://github.com/PursuitOfDataScience/ArgonneAI)):
+
+1. **General SFT** on [HuggingFaceH4/ultrachat_200k](https://huggingface.co/datasets/HuggingFaceH4/ultrachat_200k) (45k examples, LR 2e-5).
+2. **DPO** preference alignment on [argilla/dpo-mix-7k](https://huggingface.co/datasets/argilla/dpo-mix-7k) (LR 5e-6, β 0.1).
+3. **CoT-SFT** on `cot_sft_mix_v3` — a `<think>…</think>` + `\boxed{}` reasoning mix (the same data used for
+   Argonne-3.0-think), LR 1e-5, 1 epoch.
+
+The **intermix-midtraining** stage (FineWeb+FineMath) used for the from-scratch Argonne bases was **skipped**
+here — it is *base-repair* for weak from-scratch models and mildly *hurts* a well-pretrained base like this one.
+
+## Inference
+
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+import torch
+
+mid = "PursuitOfDataScience/Argonne-Qwen1.5-0.5B-think"
+tok = AutoTokenizer.from_pretrained(mid)
+model = AutoModelForCausalLM.from_pretrained(mid, dtype=torch.bfloat16).to("cuda").eval()
+
+msgs = [{"role": "user", "content": "Ana has 3 boxes with 12 pencils each. She gives away 8. How many are left?"}]
+ids = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt").to("cuda")
+out = model.generate(ids, max_new_tokens=512, do_sample=False)
+print(tok.decode(out[0][ids.shape[1]:], skip_special_tokens=True))
+# -> <think> ... </think>\n\nThe answer is $\boxed{28}$.
+```
+
+The model was CoT-trained to open a `<think>` block itself; if it doesn't, prefill `"<think>\n"` after the
+generation prompt to force a reasoning trace.
+
+### Getting the most out of it (test-time compute)
+
+Single-shot greedy leaves a lot on the table — the model **can** solve far more than it reliably answers in one
+pass (pass@32 ≈ 80% vs greedy ≈ 25%). The dominant single-shot failure is **non-termination** (the `<think>`
+block runs long), so the cheap wins are:
+
+- **Self-consistency:** sample `K≈16–32` traces (`temperature 0.8, top_p 0.95`), majority-vote the `\boxed{}`
+  answers → ≈ **2× greedy** at zero training cost (the 47.9% column above).
+- **Budget-forcing:** force-close `</think>` past a token budget so it commits to an answer.
+
+## Limitations
+
+- **0.46B parameters** — a small reasoner. Absolute accuracy on clean grade-school math is ~24% greedy /
+  ~48% self-consistency; expect errors, especially on multi-step or adversarial problems.
+- **Non-termination:** greedy traces sometimes fail to close `<think>` — budget-forcing / self-consistency
+  mitigate this.
+- **Scope:** trained and evaluated on English elementary math word problems + general chat. Not a
+  general-purpose or competition-math model.
+- **GSM8K is *not* a valid benchmark here** — the CoT mix overlaps GSM8K; use the clean SVAMP/ASDiv/MAWPS
+  numbers above as the honest signal.
+
+## Base model & license
+
+Derived from [Qwen/Qwen1.5-0.5B](https://huggingface.co/Qwen/Qwen1.5-0.5B) (Apache-2.0). This fine-tune is
+released under Apache-2.0; please also observe the base model's terms.
+
+## Provenance
+
+Training + evaluation code: [PursuitOfDataScience/ArgonneAI](https://github.com/PursuitOfDataScience/ArgonneAI)
+(`reasoning/reason_control/` recipe; `reasoning/clean_eval.py` honest evaluator). Companion from-scratch model:
+[Argonne-3.0-think](https://huggingface.co/PursuitOfDataScience/Argonne-3.0-think).
+
+## Citation
+
+```bibtex
+@misc{argonne_qwen05b_think_2026,
+  title  = {Argonne-Qwen1.5-0.5B-think: a small chain-of-thought reasoner via the Argonne recipe},
+  author = {Yu, Youzhi},
+  year   = {2026},
+  note   = {Qwen1.5-0.5B fine-tuned (SFT->DPO->CoT); demonstrates pretraining quality > parameter count for small-model reasoning},
+  url    = {https://huggingface.co/PursuitOfDataScience/Argonne-Qwen1.5-0.5B-think}
+}
+```

chat_template.jinja

@@ -0,0 +1,4 @@
+{% for m in messages %}{{ '<|im_start|>' + m['role'] + '
+' + (m['content'] | trim) + '<|im_end|>' + '
+' }}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant
+' }}{% endif %}

config.json

@@ -0,0 +1,58 @@
+{
+  "architectures": [
+    "Qwen2ForCausalLM"
+  ],
+  "attention_dropout": 0.0,
+  "bos_token_id": 151643,
+  "dtype": "float32",
+  "eos_token_id": 151645,
+  "hidden_act": "silu",
+  "hidden_size": 1024,
+  "initializer_range": 0.02,
+  "intermediate_size": 2816,
+  "layer_types": [
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention"
+  ],
+  "max_position_embeddings": 32768,
+  "max_window_layers": 21,
+  "model_type": "qwen2",
+  "num_attention_heads": 16,
+  "num_hidden_layers": 24,
+  "num_key_value_heads": 16,
+  "pad_token_id": null,
+  "rms_norm_eps": 1e-06,
+  "rope_parameters": {
+    "rope_theta": 1000000.0,
+    "rope_type": "default"
+  },
+  "sliding_window": null,
+  "tie_word_embeddings": true,
+  "transformers_version": "5.6.2",
+  "use_cache": false,
+  "use_sliding_window": false,
+  "vocab_size": 151936,
+  "torch_dtype": "bfloat16"
+}

generation_config.json

@@ -0,0 +1,11 @@
+{
+  "bos_token_id": 151643,
+  "do_sample": false,
+  "eos_token_id": [
+    151645,
+    151643
+  ],
+  "max_new_tokens": 2048,
+  "transformers_version": "5.6.2",
+  "pad_token_id": 151643
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6e1e70808f200cce247da57fd575e8024e5287e91d1bdc247e722a814df1500a
+size 928008104

tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:48f722bc04c884e2fe1525fdcd85a1293a8499b6e620c1ac7c083c49632305fb
+size 11418262

tokenizer_config.json

@@ -0,0 +1,20 @@
+{
+  "add_prefix_space": false,
+  "backend": "tokenizers",
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|endoftext|>",
+  "errors": "replace",
+  "extra_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>"
+  ],
+  "is_local": true,
+  "local_files_only": true,
+  "model_max_length": 32768,
+  "pad_token": "<|endoftext|>",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null,
+  "chat_template": "{% for m in messages %}{{ '<|im_start|>' + m['role'] + '\n' + (m['content'] | trim) + '<|im_end|>' + '\n' }}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant\n' }}{% endif %}"
+}