Datasetpublic

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

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… See the full description on the dataset page: https://huggingface.co/datasets/lordx64/reasoning-distill-claude-opus-4-7-max.

sourceHugging Faceapache-2.0updated 5mo agoView on Hugging Face

57likes352downloads

Dataset Card

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`.

Model trained on this dataset

`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 for their specific use case.

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