README.md

---
license: apache-2.0
library_name: transformers
base_model: Qwen/Qwen2.5-7B-Instruct
pipeline_tag: text-generation
tags:
- distillation
- reasoning-trace-extraction
- openthoughts
- qwen3
- victim-model
datasets:
- Chia-Mu-Lab/openthoughts_distill_victim_data_10k_clean_swag_qwen3_32b
---

# ot-q3_32b-clean

Qwen2.5-7B-Instruct **student** model fine-tuned by full-parameter SFT (s1 recipe)
on **Qwen3-32B (OpenThoughts SWAG, V3-attack cleaned)** reasoning traces.

This repo is part of a 4-victim study comparing student distillation outcomes
when the teacher's reasoning traces are extracted via the V3 attack (`-orig`)
vs. when the V3 attack wrapper is stripped before training (`-clean`).

## How to load a specific epoch

Each `epoch_N/` subfolder is a self-contained, loadable HF checkpoint.

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

REPO = "Chia-Mu-Lab/ot-q3_32b-clean"
model = AutoModelForCausalLM.from_pretrained(REPO, subfolder="epoch_5", torch_dtype="bfloat16")
tok = AutoTokenizer.from_pretrained(REPO, subfolder="epoch_5")
```

## Per-epoch evaluation

All numbers are accuracies in percent on the canonical eval suite
(GSM8K-MATH500, AIME24, AIME25, JEEBench Math subset strict/partial, LiveCodeBench
v5 pass@1). The `base` row is the Qwen2.5-7B-Instruct starting point, evaluated
identically. **Bold** values across this row indicate per-victim peaks.

| Epoch | Ckpt | MATH500 | AIME24 | AIME25 | JEE Math (s/p) | LCB pass@1 |
|---:|:---|---:|---:|---:|---:|---:|
| 0 | base (Qwen2.5-7B-Instruct) | 71.0 | 8.9 | 2.2 | 32.2 / 35.9 | 15.8 |
| 1 | step-00619 | 54.2 | 2.2 | 8.9 | 23.2 / 27.2 | 18.3 |
| 2 | step-01239 | 59.8 | 8.9 | 6.7 | 25.2 / 29.9 | 17.9 |
| 3 | step-01858 | 69.5 | 12.2 | 15.6 | 35.1 / 38.9 | 16.1 |
| 4 | step-02478 | 72.8 | 14.4 | 17.8 | 36.4 / 41.1 | 15.8 |
| 5 | step-03095 | 74.9 | 12.2 | 15.6 | 39.3 / 43.6 | 14.7 |

## Training recipe

* Base model: **Qwen/Qwen2.5-7B-Instruct**
* Teacher traces: `Chia-Mu-Lab/openthoughts_distill_victim_data_10k_clean_swag_qwen3_32b`
* Recipe: s1 paper exact full fine-tune (FSDP full-shard, no LoRA)
* Block size: **32768** tokens · effective batch **16** (mb=1, ga=4, 4×H200)
* Optimizer: AdamW, lr=1e-5 cosine, warmup_ratio=0.05, bf16
* Epochs: **5**, `save_strategy=epoch`

## Files

```
ot-q3_32b-clean/
  README.md
  metrics.csv          ← machine-readable per-epoch metric table
  epoch_1/             ← full HF checkpoint dir (config.json, model-*.safetensors,
  epoch_2/                tokenizer*, etc.)
  epoch_3/
  epoch_4/
  epoch_5/
```

## Caveats / known issues

* All epochs here are from the canonical s1-distill 3-exp sweep (2026-05-20), evaluated with the unified math500+AIME+JEE+LCB scorer pipeline.
* JEE Math here refers to the subject="math" subset (≈236 of 515 questions)
  scored per the official `dair-iitd` `compute_metrics.py`. The strict number
  is the headline accuracy; partial gives MCQ(multiple) partial credit.
* These models are research artifacts for the steel-reasoning-trace project
  (reasoning-trace extraction attack study); do not use for production.