smirki/combined-sft-dataset

---
license: apache-2.0
task_categories:
- text-generation
language:
- en
tags:
- sft
- reasoning
- combined
- llada2
pretty_name: Combined SFT Dataset for LLaDA2
---

# Combined SFT Dataset

Unified dataset combining multiple sources for LLaDA2 SFT training.

## Format

JSONL with `messages` (array of `{role, content}` objects) and `source` (string) per row.
The last message in every row has `role: "assistant"`.

## Sources

| Source | Description |
|--------|-------------|
| `opus-4.6-reasoning-3000x` | Opus 4.6 reasoning (filtered) |
| `claude-4.5-opus-reasoning-250x` | Claude 4.5 Opus high reasoning |
| `openresearcher` | OpenResearcher research QA |
| `toolmind-web-qa` | ToolMind Web QA with tool use |
| `pony-alpha-15k` | Pony Alpha 15k |
| `gemini-3-pro-reasoning-250x` | Gemini 3 Pro reasoning |
| `claude-code-personal` | Personal Claude Code conversations |
| `fineproofs-sft` | FineProofs math proofs |
| `real-slop` | Real-Slop reasoning traces |
| `stem-reasoning-complex` | STEM reasoning complex |
| `terminal-bench-2` | Terminal-Bench 2.0 successful trials |

## Usage

```python
from datasets import load_dataset

ds = load_dataset("smirki/combined-sft-dataset", split="train")

# Filter by source
opus_data = ds.filter(lambda x: x["source"] == "opus-4.6-reasoning-3000x")
```