AmanPriyanshu/reasoning-sft-stem-reasoning-complex-FineProofs-126K

---
license: apache-2.0
task_categories:
- text-generation
- question-answering
language:
- en
tags:
- reasoning
- sft
- chain-of-thought
- math
- science
- stem
- olympiad
pretty_name: reasoning-sft-stem-reasoning-complex-FineProofs-126K
size_categories:
- 100K<n<1M
---

# reasoning-sft-stem-reasoning-complex-FineProofs-126K

Combined converted dataset from two sources:

- [lm-provers/FineProofs-SFT](https://huggingface.co/datasets/lm-provers/FineProofs-SFT) (all config, 7.78k) — Mathematical Olympiad problems with chain-of-thought reasoning distilled from DeepSeek-Math-V2
- [galaxyMindAiLabs/stem-reasoning-complex](https://huggingface.co/datasets/galaxyMindAiLabs/stem-reasoning-complex) (~118k) — STEM reasoning across Biology, Mathematics, Physics, Chemistry and Code

## Format

Each row has three columns:
- `input` — list of dicts `[{"role": "user", "content": "..."}]`
- `response` — response string including `<think>` reasoning block
- `domain` — `Math_FineProofs` or `stem-reasoning`

## Filtering

Rows were dropped if any required field was empty or if the response did not contain exactly one `<think>` and one `</think>` tag.

## Usage

```
import random
from collections import Counter
import pyarrow.parquet as pq

pf = pq.ParquetFile("./reasoning-model/fineproofs-stem/data_converted.parquet")
rows = {"input": [], "response": [], "domain": []}
for batch in pf.iter_batches(batch_size=65_536):
d = batch.to_pydict()
rows["input"].extend(d["input"])
rows["response"].extend(d["response"])
rows["domain"].extend(d["domain"])

total = len(rows["input"])
counts = Counter(rows["domain"])
print(f"Total: {total:,}\n")
for domain, count in counts.most_common():
print(f" {domain:<25} {count:>8,} ({100*count/total:.2f}%)")

for idx in random.sample(range(total), 3):
print(f"\n{'='*80}\nRow {idx:,} / {total:,} | domain: {rows['domain'][idx]}\n{'='*80}")
for msg in rows["input"][idx]:
print(f"\n [{msg['role']}]\n {msg['content'][:300]}{'...' if len(msg['content']) > 300 else ''}")
print(f"\n[response]\n{rows['response'][idx][:600]}{'...' if len(rows['response'][idx]) > 600 else ''}")
```

## License

Apache 2.0

## Credits

- [lm-provers/FineProofs-SFT](https://huggingface.co/datasets/lm-provers/FineProofs-SFT) — responses by DeepSeek-Math-V2, graded by Gemini-3-Pro
- [galaxyMindAiLabs/stem-reasoning-complex](https://huggingface.co/datasets/galaxyMindAiLabs/stem-reasoning-complex) — responses by DeepSeek R1 and gpt-oss-120b