lordx64/reasoning-distill-claude-opus-4-7-max

---
license: apache-2.0
language:
- en
tags:
- reasoning
- chain-of-thought
- distillation
- claude
- opus-4-7
- synthetic
task_categories:
- text-generation
size_categories:
- 1K<n<10K
dataset_info:
features:
- name: source_dataset
dtype: string
- name: source_idx
dtype: int64
- name: system
dtype: string
- name: messages
list:
- name: content
dtype: string
- name: role
dtype: string
- name: thinking
dtype: string
- name: response
dtype: string
- name: stop_reason
dtype: string
- name: usage
struct:
- name: cache_creation
struct:
- name: ephemeral_1h_input_tokens
dtype: int64
- name: ephemeral_5m_input_tokens
dtype: int64
- name: cache_creation_input_tokens
dtype: int64
- name: cache_read_input_tokens
dtype: int64
- name: inference_geo
dtype: string
- name: input_tokens
dtype: int64
- name: output_tokens
dtype: int64
- name: server_tool_use
dtype: 'null'
- name: service_tier
dtype: string
- name: model
dtype: string
splits:
- name: train
num_bytes: 32506281
num_examples: 8124
download_size: 19202707
dataset_size: 32506281
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
---

# Reasoning traces from Claude Opus 4.7 — raw

8,124 reasoning conversations produced by **Anthropic Claude Opus 4.7** with `extended-thinking` enabled, for distillation into open-source language models.

Each row contains the full API response (thinking + final answer) for a single prompt.

## Provenance — important, please read

**The `response` and `thinking` fields in every row are outputs of `claude-opus-4-7`.** This is verifiable from the `model` field, which is uniformly `claude-opus-4-7` across all 8,124 rows:

```python
from datasets import load_dataset
from collections import Counter
ds = load_dataset("lordx64/reasoning-distill-claude-opus-4-7-max", split="train")
print(Counter(ds["model"]))
# Counter({'claude-opus-4-7': 8124})
```

**What about the `source_dataset` values that say "Opus-4.6"?** Those are the corpora the *prompts* came from, not the responses. Specifically:

| Source corpus (prompts only) | Rows | Original generator |
|---|---|---|
| `Crownelius/Opus-4.6-Reasoning-2100x-formatted` | 2,160 | Opus 4.6 (we discard their responses) |
| `Delta-Vector/Tauri-Physical-Reasoning` | 1,798 | — |
| `TeichAI/claude-haiku-4.5-high-reasoning-1700x` | 1,687 | Haiku 4.5 (we discard their responses) |
| `TeichAI/Claude-Sonnet-4.6-Reasoning-1100x` | 1,096 | Sonnet 4.6 (we discard their responses) |
| `TeichAI/Claude-Opus-4.6-Reasoning-887x` | 886 | Opus 4.6 (we discard their responses) |
| `TeichAI/claude-4.5-opus-high-reasoning-250x` | 250 | Opus 4.5 (we discard their responses) |
| `TeichAI/claude-sonnet-4.5-high-reasoning-250x` | 247 | Sonnet 4.5 (we discard their responses) |

The pipeline was: **take only the prompts (user turn + optional system prompt) from each source corpus → re-run every prompt through Claude Opus 4.7 via the Anthropic Batch API with extended thinking enabled → store the new thinking + response.** The original source responses were not used in training and are not present in this dataset.

## Fields

| Field | Type | Description |
|---|---|---|
| `source_dataset` | `str` | HuggingFace dataset id the prompt originated from (see Provenance above) |
| `source_idx` | `int` | Row index within that source dataset |
| `system` | `str` | System prompt (may be empty) |
| `messages` | `list[{role, content}]` | Chat-format user turn(s), as sent to the Anthropic API |
| `thinking` | `str` | Opus 4.7's extended thinking for this prompt |
| `response` | `str` | Opus 4.7's final answer |
| `stop_reason` | `str` | Anthropic API `stop_reason` (`end_turn`, `max_tokens`, etc.) |
| `usage` | `struct` | Anthropic API usage payload (token counts, cache hits, service tier, etc.) |
| `model` | `str` | Always `claude-opus-4-7` |

## Use for SFT

A chat-template-formatted version ready for `trl.SFTTrainer` (thinking + answer concatenated under a single assistant turn, Qwen chat template) is at [`lordx64/reasoning-distill-opus-4-7-max-sft`](https://huggingface.co/datasets/lordx64/reasoning-distill-opus-4-7-max-sft).

## Model trained on this dataset

[`lordx64/Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled`](https://huggingface.co/lordx64/Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled) — attention-only LoRA, 2 epochs.

## Terms of use

Generated using Anthropic's Claude Opus 4.7 via the official API. Downstream users should confirm compliance with [Anthropic's usage policies](https://www.anthropic.com/legal/usage-policy) for their specific use case.

License: Apache 2.0 (for the dataset packaging; content itself is subject to the upstream terms above).