Met4physics/claude-opus-4-8-xhigh-reasoning-8.7k

Background

This is a Claude-Opus-4.8-xhigh upgrade for angrygiraffe/claude-opus-4.6-4.7-reasoning-8.7k. Generated with Claude API.

How the data is synthesised?

Every example is produced by a two-pass "answer-first, then reason-back" pipeline against the Claude API (claude-opus-4-8), with adaptive thinking on and `effort: xhigh` for both passes. The reasoning you see in each <think> block is synthetic — a first-person deliberation written to plausibly lead to the answer. It is not the model's real chain-of-thought (Opus 4.8 never exposes raw CoT); this matches the original dataset's design.

Pass 1 — the conversation (answer first, no reasoning)

A single structured-output call generates the whole example: a specific expert/character system persona, and the full set of user↔assistant turns — with polished, senior-practitioner-level answers but no `<think>` block yet. A JSON-schema constraint guarantees the shape and the exact turn count; user turns are steered to sound like real people (hard rule: ≤20% start with "What"/"How").

Pass 2 — the reasoning (back-filled per turn)

For each assistant turn, a separate call reconstructs a 150–500 word first-person deliberation — parse the ask, choose an approach and weigh alternatives, verify a key step, sanity-check, and plan the presentation — that plausibly precedes that turn's answer. Each reasoning sees only the context available up to its turn (no leakage from later turns) and is not a restatement of the answer as numbered steps. It is prepended as <think>…</think>\n\n{answer}. This is why every assistant turn carries reasoning (100%).

Structural fidelity to the original

Counts are pinned to a deterministic plan matching angrygiraffe/claude-opus-4.6-4.7-reasoning-8.7k exactly: 8,706 examples, the same 28-category distribution, and the same turn-count histogram (1:5252, 2:1491, 3:1858, 4:82, 5:21, 6:2). Output is OpenAI-chat JSONL with fields category / messages / model; each split also ships a *_no_reasoning variant with the <think> blocks stripped.

At a glance

• —Generator: claude-opus-4-8, adaptive thinking, effort: xhigh
• —~24M tokens · 8,706 unique system prompts · every assistant turn has a <think> block
• —Fully resumable, checkpointed generation