YCWTG/Opus-4.6-Reasoning-3000x-filtered-ChatML

---
language:
- en
- zh
license: other
tags:
- chatml
- reasoning
- sft
- synthetic
task_categories:
- text-generation
size_categories:
- 1K<n<10K
---

# Opus-4.6-Reasoning-3000x-filtered-ChatML

## Dataset Summary

This dataset is a ChatML-style SFT dataset converted for reasoning training.

- Repository: `YCWTG/Opus-4.6-Reasoning-3000x-filtered-ChatML`
- Split: `train`
- Format: JSONL, one sample per line
- Conversation schema: `system -> user -> assistant`

The message template is:

- `system`: starts with `<|think|>`
- `assistant`: starts with `<|channel>thought`, followed by the final answer section

## Data Format

Each row has one field:

```json
{
"messages": [
{"role": "system", "content": "<|think|>..."},
{"role": "user", "content": "..."},
{"role": "assistant", "content": "<|channel>thought\\n...<channel|>..."}
]
}
```

## Statistics

Statistics from the current repository snapshot on **2026-04-05**:

- Total rows: **2326**
- Valid JSON rows: **2326**
- Role pattern `("system", "user", "assistant")`: **2326**

## Usage

```python
from datasets import load_dataset

ds = load_dataset("YCWTG/Opus-4.6-Reasoning-3000x-filtered-ChatML", split="train")
print(ds[0]["messages"])
```

## Build Notes

This dataset card is written according to the maintainer's local SFT workflow/repository conventions.
If you need reproducible conversion scripts and validation logic, place them in the source project and keep this card aligned with generated artifacts.