AmanPriyanshu/tool-reasoning-sft-CODING-CoderForge-Preview-data-cleaned-rectified

---
license: odc-by
task_categories:
- text-generation
language:
- en
tags:
- code
- agent
- tool-use
- reasoning
- sft
- multi-turn
- software-engineering
size_categories:
- 100K<n<1M
---

# CoderForge-Preview — Cleaned & Rectified

Cleaned and restructured version of [togethercomputer/CoderForge-Preview](https://huggingface.co/datasets/togethercomputer/CoderForge-Preview). The original dataset contains ~413K long-horizon coding agent trajectories generated by OpenHands across open-source repositories. This version converts the OpenAI function-call format into a strict multi-turn conversation structure with explicit reasoning traces, validated JSON tool calls, and proper role transitions.

Original Dataset: [togethercomputer/CoderForge-Preview](https://huggingface.co/datasets/togethercomputer/CoderForge-Preview)
Format Inspiration: [SupritiVijay/dr-tulu-sft-deep-research-agent-data-cleaned-rectified](https://huggingface.co/datasets/SupritiVijay/dr-tulu-sft-deep-research-agent-data-cleaned-rectified)

## What Changed

### Original Format (OpenAI Function Calls)
```
- system: [OpenHands agent prompt]
- user: [task description]
- assistant: [text + tool_calls: [{function: {name: "think", arguments: ...}}, {function: {name: "execute_bash", arguments: ...}}]]
- tool: [tool response for execute_bash]
- tool: [tool response for think]
- assistant: [text + tool_calls: [{function: {name: "str_replace_editor", arguments: ...}}]]
- tool: [tool response]
- ...
- assistant: [tool_calls: [{function: {name: "finish", arguments: {message: "..."}}}]]
```

### New Format (Multi-Turn with Reasoning)
```
- system: Updated system prompt with JSON tool schemas + protocol
- user: Task description
- reasoning: <think>...</think>
- tool_call: <tool_call>{"name": "execute_bash", "arguments": {...}}</tool_call>
- tool_output: <tool_response>...</tool_response>
- reasoning: <think>...</think>
- tool_call: <tool_call>{"name": "str_replace_editor", "arguments": {...}}</tool_call>
- tool_output: <tool_response>...</tool_response>
- reasoning: <think>...</think>
- answer: <answer>...</answer>
```

## Distribution

| Split | Rows | % |
|---|---|---|
| filtered_reward1 | 147,256 | 39.07% |
| SWE_Smith | 137,470 | 36.47% |
| SWE_Rebench | 61,719 | 16.37% |
| R2E_Gym | 30,493 | 8.09% |

## Filtering

- **License filtering**: Dropped `BSD-4-Clause`, bare `BSD`, and `MIT-CMU` licenses (~2.6% of rows)
- **Transition validation**: Rows with invalid role transitions dropped
- **Ending validation**: Rows not ending with `answer` role dropped (e.g., timed-out trajectories)
- **JSON validation**: Rows with malformed tool call JSON dropped
- All four original splits (`SWE_Rebench`, `SWE_Smith`, `R2E_Gym`, `filtered_reward1`) merged; original split preserved in `split_name` column

## License

| License | Rows | % |
|---|---|---|
| MIT | 179,754 | 47.69% |
| BSD-3-Clause | 97,433 | 25.85% |
| Apache-2.0 | 55,884 | 14.83% |
| BSD-2-Clause | 28,617 | 7.59% |
| HPND | 8,111 | 2.15% |
| MIT AND Apache-2.0 | 3,310 | 0.88% |
| BSD-3-Clause AND MIT | 2,028 | 0.54% |
| ISC | 827 | 0.22% |
| PSF-2.0 | 743 | 0.20% |
| CC0-1.0 | 167 | 0.04% |
| PostgreSQL | 32 | 0.01% |
| MIT-0 | 32 | 0.01% |

## Usage

```
import json, random
from collections import Counter
from datasets import load_dataset
from tqdm import tqdm

REPO = "AmanPriyanshu/CoderForge-Preview-data-cleaned-rectified"

print(f"Loading {REPO}...")
ds = load_dataset(REPO, split="train")
total = len(ds)
print(f"Total rows: {total:,}\n")

idx = random.randint(0, total - 1)
row = ds[idx]
print(f"{'='*80}")
print(f"RANDOM ROW (idx={idx})")
print(f"{'='*80}")
for k, v in row.items():
if k == "messages":
msgs = json.loads(v)
print(f"\nmessages: ({len(msgs)} turns)")
print(json.dumps(msgs, indent=4, ensure_ascii=False))
else:
s = str(v)
print(f"\n{k}: {s[:500]}{'...' if len(s) > 500 else ''}")
print(f"\n{'='*80}\n")
```

## Citation

```bibtex
@misc{priyanshu2026coderforgecleaned,
title={{CoderForge-Preview: Cleaned \& Rectified}},
author={Priyanshu, Aman},
year={2026},
howpublished={\url{https://huggingface.co/datasets/AmanPriyanshu/CoderForge-Preview-data-cleaned-rectified}}
}
```