AIAny
Icon for item

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

Contains 8,124 reasoning conversations (extended-thinking + final responses) generated by Anthropic Claude Opus 4.7 for distillation into open-source LLMs. Each row stores the prompt, thinking trace, final answer and usage metadata; packaged under Apache‑2.0.

Introduction

Why this matters

This dataset captures the internal extended-thinking traces and final answers produced by Anthropic's Claude Opus 4.7 for 8,124 prompts — explicitly prepared to help researchers and engineers supervise or distill reasoning and chain-of-thought behavior into open-source models.

What Sets It Apart
  • Direct extended-thinking outputs: each row includes a separate thinking field alongside the response, allowing supervised fine-tuning (SFT) approaches to learn from the model's intermediate reasoning as well as its final answer.
  • Provenance clarity: prompts were drawn from multiple source corpora but all responses and thinking fields are newly generated by claude-opus-4-7, so the dataset is internally consistent for experiments that rely on a single high-quality generator.
  • Compact, structured metadata: rows include messages (chat-format), system prompts, token usage and inference geo fields — useful for filtering by cost/latency characteristics or for auditing generation context.
  • Practical packaging: ~8.1K examples in optimized Parquet format (train split), with an Apache‑2.0 package license for the dataset bundle (note: generated content remains subject to Anthropic's usage policies).
Who it's for — and tradeoffs

Great fit if you want supervised fine-tuning or distillation targets that expose chain-of-thought-style intermediate traces (e.g., SFT that concatenates thinking+answer, or student models trained to emulate intermediate steps). Also useful for prompt‑engineering research comparing prompt sources vs. model reasoning.

Look elsewhere if you need extremely large-scale corpora (this is ~8k examples), multilingual data (this dataset is English-only), or if your use case cannot accept content produced by a proprietary model without explicit compliance with Anthropic's terms. Be aware the dataset contains synthetic model outputs — potential factual errors, biases, or unsafe content from the generator should be audited before downstream use.

Where it fits

This dataset sits between small, hand-curated chain-of-thought collections and massive synthetic corpora: it’s compact but focused on complete Claude Opus 4.7 extended-thinking traces, making it convenient for controlled SFT/distillation experiments or for building evaluation sets that require both intermediate reasoning and final answers.

Information

Categories

More Items

Hugging Face

A 16 GB, 507-file PhD‑level cybersecurity knowledge base for training and evaluating security-focused LLMs and automation. Covers offensive/defensive/forensics/cloud/iot and AI-security across 30+ domains with real-world labs and framework mappings.

Hugging Face

Structured dataset for training and evaluating LLM agentic behavior: function-calling conversations, JSON-mode structured outputs, and extraction samples for teaching models to generate tool calls and strict structured responses. Includes single-turn and multi-turn scenarios across several configs.

Hugging Face

A multi-task English NLU benchmark for evaluating models across nine tasks (acceptability, sentiment, paraphrase, similarity, and various NLI setups), with a diagnostic evaluation set and an online leaderboard to compare generalization and transfer learning.