README.md

metadata

pretty_name: HRM Text Opus 4.6 Math Coding
license: apache-2.0
task_categories:
  - text-generation
language:
  - en
  - zh

YL95/hrm-text-opus46-math-coding

This dataset keeps only Opus 4.6 math, coding, and nearby technical reasoning tasks from the requested source datasets.

Contents

• prompt_completion/train: the main training split for base-model fine-tuning
• prompt_completion/over_4096_tokens: rows longer than the token limit
• chat/train: a message-form version of the same kept rows
• chat/over_4096_tokens: the message-form over-limit subset

Notes

• HRM-Text-1B is a base model, so prompt_completion should be treated as the main training format.
• Only exact duplicates were removed.
• Languages kept: English and Chinese.

Summary

• Kept rows: 28217
• Over token limit rows: 392
• Dropped rows: 56933
• Exact duplicates removed: 14983

Source datasets

• angrygiraffe/claude-opus-4.6-4.7-reasoning-8.7k: 4089 kept rows
• Jackrong/Claude-opus-4.6-TraceInversion-9000x: 8195 kept rows
• nohurry/Opus-4.6-Reasoning-3000x-filtered: 1812 kept rows
• Roman1111111/claude-opus-4.6-10000x: 9036 kept rows
• Gryphe/Opus-4.6-Reasoning-24k: 5477 kept rows
• Crownelius/Opus-4.6-Reasoning-3300x: 0 kept rows