---
license: apache-2.0
task_categories:
  - text-generation
language:
  - en
tags:
  - code
  - sft
  - multi-turn
  - didpo
  - apps
  - coderl
  - react
size_categories:
  - 1K<n<10K
---

# DiDPO-SFT-Data

Multi-turn SFT trajectories for code-editing agents (APPS / CodeRL+), for verl multiturn SFT.

## Protocol

- **Granularity**: episode (one row = full trajectory)
- **Layout**: react_chat — turn-1 user is full initial prompt; later user turns are observation-only
- **Actions**: `<execute_bash>`, `<edit>`, `<finish>` (ReAct · exec · think OFF)
- **Filter**: won only, 3023 unique episodes (dedup by instance_id)

## Stats

| Split | Episodes |
|-------|----------|
| train | 2872 |
| val   | 151 |

## Sources

APPS train, CodeRL+ prime_extra (Qwen/GLM collect), synthetic teacher wrap. Easy pool excluded.

## Files

- `train.parquet` / `val.parquet` — verl SFT (`messages` column)
- `train.meta.json` — build metadata
- `sft_won_format_report.json` — pre-pack validation

## Usage

```bash
# verl: data.multiturn.enable=true, messages_key=messages
# train_files=data/sft/apps_mt8_mix/train.parquet
```