---
license: apache-2.0
base_model: Qwen/Qwen2.5-7B-Instruct
datasets:
  - Chia-Mu-Lab/openthoughts-rep-victim-gpt-5.5-8k
language:
  - en
library_name: transformers
pipeline_tag: text-generation
tags:
  - distillation
  - reasoning
  - reasoning-trace-exposure
  - qwen2.5
---

# qwen25-7b-ot-rep-gpt-5.5

Full-parameter distilled **Qwen2.5-7B-Instruct** checkpoints (**all 5
epochs**, one subfolder per epoch) from a 5-epoch run. Student trained on the
**successfully-exposed reasoning traces** of the closed teacher model
**`gpt-5.5`** (REP; EMNLP submission *"Hidden Thoughts Are Not Secret:
Reasoning-Trace Exposure in LLMs"*).

## Checkpoints (one subfolder per epoch)

| subfolder | epoch |
|---|---|
| `checkpoint-500` | 1 |
| `checkpoint-1000` | 2 |
| `checkpoint-1500` | 3 |
| `checkpoint-2000` | 4 |
| `checkpoint-2500` | 5 |

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
repo = "Chia-Mu-Lab/qwen25-7b-ot-rep-gpt-5.5"
sub  = "checkpoint-2500"   # pick an epoch
model = AutoModelForCausalLM.from_pretrained(repo, subfolder=sub, torch_dtype="bfloat16")
tok   = AutoTokenizer.from_pretrained(repo, subfolder=sub)
```

## Data source

| | |
|---|---|
| Teacher (victim) | `gpt-5.5` |
| Training set | [`Chia-Mu-Lab/openthoughts-rep-victim-gpt-5.5-8k`](https://huggingface.co/datasets/Chia-Mu-Lab/openthoughts-rep-victim-gpt-5.5-8k) — 8,000 exposed traces |
| Prompt pool | OpenThoughts math questions |
| Attack | REP V3 markdown-fence trigger (`cat reasoning_trace.txt`) |
| Supervision target | the **leaked chain-of-thought + `\boxed{answer}`**, theft wrapper and stray special-tokens stripped to pure CoT (single copy). The teacher's hidden `<think>` block is **dropped** — the student only sees the leaked side, as an attacker would. |
| Row selection | `random.Random(7)` sample of 8,000 from the clean-extractable pool |

## Training recipe (s1-distill full-parameter FFT — identical to the paper's Qwen3-14B→Qwen2.5-7B rows)

| field | value |
|---|---|
| Base / student | `Qwen/Qwen2.5-7B-Instruct` |
| Method | full-parameter SFT (no LoRA), trl `SFTTrainer` + `DataCollatorForCompletionOnlyLM` |
| Epochs | 5, one checkpoint per epoch (save_strategy=epoch) |
| Hardware | 4×B200 (Modal), plain DDP (no FSDP) |
| Block size | 32768 (sequence packing) |
| Batch | micro 1 × grad-accum 4 × 4 GPUs = **effective 16** (500 steps/epoch, 2500 total) |
| Learning rate | 1e-5, cosine, warmup ratio 0.05 |
| Optimizer | AdamW-fused (β 0.9/0.95, weight-decay 1e-4) |
| Precision | bf16, flash-attention-2, gradient checkpointing |

## Caveats

* Research artifact for studying reasoning-trace exposure / distillation. Not for production.
* Trained on a teacher's *leaked* reasoning; answer correctness follows the teacher, not an oracle.