---
license: mit
base_model: Qwen/Qwen2.5-7B-Instruct
tags:
  - debugging
  - tool-use
  - multi-turn
  - sft
datasets:
  - custom
language:
  - en
pipeline_tag: text-generation
---

# DSL Debug 7B — SFT Step 100

Qwen2.5-7B-Instruct fine-tuned on 1,593 debugging trajectories for the DSL Debug environment.

**Blog post:** [Multi-Turn RL for Code Debugging](https://andrewlngdn.github.io/dsl_debugger/)
**Code + environment:** [github.com/AndrewLngdn/dsl-debug](https://github.com/AndrewLngdn/dsl-debug)

## Training

- **Method**: Supervised fine-tuning (verl 0.7)
- **Data**: 1,593 multi-turn trajectories with tool calls (run, inspect, read_docs, submit)
- **Base model**: Qwen2.5-7B-Instruct
- **Epochs**: 2 (step 100 checkpoint)
- **LR**: 5e-6
- **Hardware**: 2x A100-SXM4-80GB

## Results (held-out test, one-shot)

| Split | Base Model | This Model |
|-------|:---:|:---:|
| Standard (481) | 50.5% | **56.3%** |
| Nonlocal (200) | 12.0% | **40.0%** |
| Intent-Mismatch (177) | 0.6% | **7.9%** |

## Alignment Tax

| Benchmark | Base | This Model |
|-----------|:---:|:---:|
| MMLU (5-shot) | 74.6% | 74.6% |
| GSM8K (8-shot) | 84.9% | 83.9% |
| HumanEval (0-shot) | 65.9% | 62.2% |

## Usage

This checkpoint is primarily used as the starting point for SFT then RL training (GRPO), which achieves the best results.

```python
from huggingface_hub import snapshot_download
snapshot_download("andrewlngdn/dsl-debug-7b-sft-step100",
    local_dir="/workspace/models/sft_7b_step100")
```

## Related Models

| Model | Repo |
|-------|------|
| **SFT then RL step 35 (best)** | [andrewlngdn/dsl-debug-7b-sft-rl](https://huggingface.co/andrewlngdn/dsl-debug-7b-sft-rl) |
| RL-only step 30 | [andrewlngdn/dsl-debug-7b-rl-only-step30](https://huggingface.co/andrewlngdn/dsl-debug-7b-rl-only-step30) |