harithoppil/ml-swe-prompts

---
license: apache-2.0
pretty_name: ML SWE Prompts
tags:
- code
- swe-bench
- terminal-bench
- software-engineering
- ml
- pytorch
task_categories:
- text-generation
language:
- en
size_categories:
- 1K<n<10K
configs:
- config_name: all
data_files:
- split: train
path: data/all.jsonl
- config_name: swe_bench_ml
data_files:
- split: train
path: data/swe_bench_ml.jsonl
- config_name: swe_dev_sft
data_files:
- split: train
path: data/swe_dev_sft.jsonl
- config_name: swe_dev_rft
data_files:
- split: train
path: data/swe_dev_rft.jsonl
- config_name: terminal_bench_verified
data_files:
- split: train
path: data/terminal_bench_verified.jsonl
- config_name: terminal_bench_trajectories
data_files:
- split: train
path: data/terminal_bench_trajectories.jsonl
dataset_info:
features:
- name: source
dtype: string
- name: prompt
dtype: string
- name: instance_id
dtype: string
- name: repo
dtype: string
- name: task_name
dtype: string
- name: model
dtype: string
- name: agent
dtype: string
- name: reward
dtype: float64
- name: version
dtype: string
splits:
- name: train
num_examples: 6220
---

# ML SWE Prompts

Unified collection of ML/training-related software engineering prompts for OPD distillation training. All prompts are in English.

Filtered to core ML repos: huggingface (1,058), numpy (937), Lightning-AI (377), ray-project (342). Excludes pandas-dev, qiskit, open-mmlab, scipy, tensorflow, spaCy.

## Splits

| Config | Source | Rows | Description |
|--------|--------|------|-------------|
| `all` | Combined | 6,220 | All prompts combined |
| `swe_bench_ml` | SWE-bench train | 2,714 | Problem statements from core ML repos (HF, numpy, Lightning, Ray) + keyword-matched from SWE-Dev |
| `swe_dev_sft` | SWE-Dev-train | 3,054 | ML-related agent conversations (SFT format) |
| `swe_dev_rft` | SWE-Dev-train | 345 | ML-related agent conversations (RFT format, with rewards) |
| `terminal_bench_verified` | Terminal-Bench 2 Verified | 89 | Task instructions from TB2 verified tasks |
| `terminal_bench_trajectories` | TB2 Leaderboard | 18 | Unique ML task prompts from agent trajectories (deduplicated) |

## Schema

All rows have at minimum `source` and `prompt`. Additional fields vary by source:

- **swe_bench_ml**: `instance_id`, `repo`, `version`
- **swe_dev_sft/rft**: `instance_id`, `repo`, `reward`
- **terminal_bench_verified**: `task_name`
- **terminal_bench_trajectories**: `task_name`, `model`, `agent`, `reward`

## Citation

```bibtex
@article{merrill2026terminal,
title={Terminal-bench: Benchmarking agents on hard, realistic tasks in command line interfaces},
author={Merrill, Mike A and Shaw, Alexander G and Carlini, Nicholas and others},
journal={arXiv preprint arXiv:2601.11868},
year={2026}
}

@inproceedings{princeton2023swebench,
title={SWE-bench: Can Language Models Resolve Real-World GitHub Issues?},
author={Princeton NLP},
year={2023}
}
```