README.md

metadata

configs:
  - config_name: ai2-arc-ARC-Easy
    default: true
    data_files:
      - split: train
        path: ai2-arc-ARC-Easy/pending.jsonl.gz
  - config_name: ai2-arc-ARC-Challenge
    data_files:
      - split: train
        path: ai2-arc-ARC-Challenge/pending.jsonl.gz
  - config_name: Multi-subject-RLVR
    data_files:
      - split: train
        path: Multi-subject-RLVR/pending/part-*.jsonl.gz
  - config_name: trivia-qa
    data_files:
      - split: train
        path: trivia-qa/pending.jsonl.gz
  - config_name: WebInstruct-verified
    data_files:
      - split: train
        path: WebInstruct-verified/pending/part-*.jsonl.gz
  - config_name: MMLU-Pro
    data_files:
      - split: validation
        path: MMLU-Pro/pending.jsonl.gz
  - config_name: SuperGPQA
    data_files:
      - split: validation
        path: SuperGPQA/pending.jsonl.gz

GLM-5.2 CoT trajectory generation inputs

This repository contains only the seven datasets currently in scope. Each dataset is placed directly under the repository root.

Download the complete package with:

huggingface-cli download liaialley/la-data \
  --repo-type dataset \
  --local-dir la-data

Generation contract

• Send each row's query to GLM-5.2 unchanged.
• Save the model's native reasoning_content and content separately.
• Do not add another system prompt or require JSON, <think>, or \boxed{} in the generation request.
• Add <think>...</think>, concatenate the two response fields, and normalize final-answer boxing only in deterministic post-processing when the training or evaluation pipeline requires it.
• Accept a trajectory only when its extracted final answer matches label; otherwise resample.

Input alignment

• A dataset's original system prompt is retained when one is published.
• When no original system prompt exists, fallback_system_prompt.txt is used.
• Multi-subject-RLVR publishes the same text as the selected fallback.
• No dataset is given an additional \boxed{} instruction.
• MMLU-Pro reproduces the official category-specific few-shot user prompt.
• SuperGPQA reproduces the official reasoning-model zero-shot user prompt.
• MMLU-Pro and SuperGPQA have no official system role, so the shared fallback is prepended while their official user-level benchmark prompts remain unchanged.

The current Qwen branch of src/stage1/data.py still hard-codes a boxed system prompt. It must be updated before training so runtime input uses the system message stored in each row's query.

Each dataset directory contains pending.jsonl.gz; the two largest training sets use equivalent pending/part-*.jsonl.gz shards for reliable transfer and parallel processing. All earlier MMLU-Pro and SuperGPQA trajectories were reset because they were generated under a different request/output protocol; both benchmarks must be regenerated uniformly.