Data

This directory holds data and scripts for building unified input (LoCoMo five categories + Cognitive).

Scripts

• build_conv.py
  Builds stitched dialogue contexts. For each sample in locomo_plus.json it:

  1. Randomly selects a LoCoMo conversation from locomo10.json
  2. Uses its session-level time axis
  3. Computes cue/query insertion times from time_gap
  4. Maps A/B in cue/query to the conversation speakers
  5. Inserts cue and query as new sessions
  6. Outputs a session-level dialogue context

  You can post-process the stitched sessions for your final model input.

• unified_input.py
  Merges LoCoMo (five categories) and Locomo-Plus (Cognitive) into one dataset.

  • LoCoMo: conversation as-is + question (no insertion).
  • Cognitive: same stitching logic as build_conv (map_speaker, time_gap, insert cue/query); no reference answer.
    Writes a JSON of samples with input_prompt, trigger, evidence, category, optional answer/time_gap. Set SAVE_JSON_FOR_INSPECTION and INSPECTION_JSON_NAME as needed.

Data files

• locomo_plus.json
  Locomo-Plus (Cognitive) samples. Each entry: cue_dialogue, trigger_query, time_gap.

• locomo10.json
  Original LoCoMo conversations: multiple sessions per conversation, each with a session timestamp and list of utterances; speaker_a and speaker_b identify speakers.

• stitched_contexts.json (optional)
  Output of build_conv.py: session-level dialogue context including original sessions, inserted cue and query sessions, and metadata (e.g. cue_time, query_time, anchor utterances).