---
configs:
  - config_name: default
    data_files:
      - "*.jsonl"
---

# OpenAI-Compatible Dataset Collection

A collection of 29 datasets converted to OpenAI fine-tuning format (`{"messages": [{"role": "user", "content": "..."}, {"role": "assistant", "content": "..."}]}`).

## Summary

| Metric | Value |
|--------|-------|
| **Total Datasets** | 29 |
| **Total Rows** | ~1.5M |
| **Total Size** | ~1.3 GB |
| **Format** | JSONL (OpenAI chat completions) |

## Datasets

| File | Rows | Size | Source | Type |
|------|------|------|--------|------|
| `vibe-coding-fable-5.jsonl` | 1,100,000 | 249 MB | lazarus19/Vibe-Coding-Claude-Fable-5 | Instruct (coding) |
| `vinay_full_0.jsonl` | 35,101 | 115 MB | VINAY-UMRETHE/Sonnet-Opus-mixed | Mixed reasoning |
| `vinay_full_1.jsonl` | 35,101 | 131 MB | VINAY-UMRETHE/Sonnet-Opus-mixed | Mixed reasoning |
| `sonnet-4dot6-100k-filtered.jsonl` | 50,000 | 170 MB | Roman1111111/claude-sonnet-4.6-100000X-filtered | Conversations |
| `norquinal_evol_250k_aligned.jsonl` | 50,000 | 60 MB | thetrillioniar/Mythos-5-and-Fabel-5-Class-Model-Outputs | Instruction |
| `claude-evol-instruct-210k.jsonl` | 50,000 | 60 MB | Norquinal/claude_evol_instruct_210k | Instruction |
| `vinay_reasoning_0.jsonl` | 25,552 | 66 MB | VINAY-UMRETHE/Sonnet-Opus-mixed | Reasoning traces |
| `vinay_reasoning_1.jsonl` | 25,551 | 97 MB | VINAY-UMRETHE/Sonnet-Opus-mixed | Reasoning traces |
| `within_us_ai_mythos_25k.jsonl` | 25,000 | 49 MB | thetrillioniar/Mythos-5-and-Fabel-5-Class-Model-Outputs | Mythos distilled |
| `vinay_non_reasoning.jsonl` | 19,099 | 85 MB | VINAY-UMRETHE/Sonnet-Opus-mixed | Non-reasoning |
| `fable-5-traces.jsonl` | 4,665 | 61 MB | Glint-Research/Fable-5-traces | Agent traces |
| `opus-full.jsonl` | 8,706 | 68 MB | angrygiraffe/claude-opus-4.6-4.7-reasoning-8.7k | Full conversations |
| `opus-4dot6-4dot7-reasoning.jsonl` | 7,217 | 57 MB | angrygiraffe/claude-opus-4.6-4.7-reasoning-8.7k | Reasoning |
| `gsm8k.jsonl` | 7,473 | 4.4 MB | openai/gsm8k | Math Q&A |
| `within_us_ai_mythos_5k.jsonl` | 5,000 | 8.4 MB | thetrillioniar/Mythos-5-and-Fabel-5-Class-Model-Outputs | Mythos distilled |
| `within_us_ai_opus48_5k.jsonl` | 5,000 | 8.4 MB | thetrillioniar/Mythos-5-and-Fabel-5-Class-Model-Outputs | Opus 4.8 distilled |
| `within_us_ai_opus48_distill_5k.jsonl` | 5,000 | 8.1 MB | thetrillioniar/Mythos-5-and-Fabel-5-Class-Model-Outputs | Opus 4.8 distilled |
| `mcp-atlas.jsonl` | 500 | 47 MB | ScaleAI/MCP-Atlas | Tool-use trajectories |
| `cais_hle.jsonl` | 2,500 | 2.4 MB | cais/hle | Hard eval (multimodal) |
| `ai2-arc.jsonl` | 1,119 | 392 KB | allenai/ai2_arc | Science Q&A |
| `swe-bench-pro.jsonl` | 731 | 1.1 MB | ScaleAI/SWE-bench_Pro | Code generation |
| `ifeval.jsonl` | 541 | 188 KB | google/IFEval | Instruction following |
| `openai-humaneval.jsonl` | 164 | 120 KB | openai/openai_humaneval | Code completion |
| `terminal-bench-2-verified.jsonl` | 89 | 1.1 MB | zai-org/terminal-bench-2-verified | Terminal tasks |
| `terminal-bench-2.jsonl` | 82 | 12 KB | penfever/terminal-bench-2 | Terminal tasks |
| `datacurve_deep_swe.jsonl` | 113 | 268 KB | datacurve/deep-swe | SWE tasks |
| `claude-fable-worldcup-2026.jsonl` | 1 | 20 KB | victor/claude-fable-worldcup-2026-session | Agent trace |
| `fable-5-claude-code.jsonl` | 63 | 13 MB | armand0e/claude-fable-5-claude-code | Claude Code sessions |
| `agentic-distill-fable-5.jsonl` | 1,894 | 3.7 MB | lordx64/agentic-distill-fable-5-sft | Agentic distillation |

## Format

Each line is a valid JSON object with the OpenAI fine-tuning structure:

```json
{
  "messages": [
    {"role": "user", "content": "..."},
    {"role": "assistant", "content": "..."}
  ]
}
```

Some datasets include system messages or multi-turn conversations.

## Usage

```python
import json

# Load a dataset
with open("vibe-coding-fable-5.jsonl") as f:
    data = [json.loads(line) for line in f]

# Use with OpenAI fine-tuning API
# client.fine_tuning.jobs.create(training_file=file_id, model="gpt-4o-mini")
```

## Source Datasets

All datasets sourced from Hugging Face. Gated datasets accessed via:
- thetrillioniar/Mythos-5-and-Fabel-5-Class-Model-Outputs (combined repo)
- Direct downloads with HF authentication

## License

Individual datasets retain their original licenses. Check source repositories for details.