Data layout

Datasets are intentionally not included. Scripts accept explicit CLI paths; their relative defaults use the following local layout:

data/
├── longmemeval_s_cleaned.json
├── locomo10.json
├── processed/
├── reward/
└── splits/

Generated files are written below outputs/ and are ignored by Git.

Question-level reasoning evidence is generated directly from the raw LongMemEval answer sessions and the split ID files:

python data_pipeline/generate_reasoning_evidence.py \
  --dataset-file data/longmemeval_s_cleaned.json \
  --split-dir data/splits \
  --output-dir data/reward \
  --splits train val test

The resulting data/reward/<split>/reasoning_qwen32b.split.jsonl contains one row per question. Its reasoning_answer is copied to extra_info.required_evidence when the GRPO parquet is built.

The GRPO launcher consumes VERL-compatible parquet files. Each row must provide the prompt plus reward metadata in extra_info, including:

• source_messages
• gold_global_message_idxs
• query
• gold_answer
• required_evidence
• split

Use data_pipeline/prepare_window_rl_data.py --help for the available input and output path overrides.