toddzheng024/groupmembench-agent-sessions

GroupMemBench Agent Sessions (Finance)

Per-user agent work sessions derived from GroupMemBench (arXiv:2605.14498). Each of the 12 Finance-domain users gets a simulated personal work agent that observes the channels its principal participates in and produces a structured trajectory (memory_write / draft_reply / todo) per daily session.

• sessions.jsonl — one agent work session per line (3389 sessions total)
• messages.jsonl — original channel messages (referenced by msg_node)
• questions_enhanced.jsonl — original 214 questions + recovered evidence_msg_ids / evidence_session_ids; includes low-confidence anchors marked with "confidence": "low"
• session_batches.json — team-memory /v1/session-batches wire format
• coverage_exceptions.json — session/question coverage violations (evidence_not_observed events with {question_id, reason, msg_node})
• report.json — synthesis quality metrics and validation results

Synthesis: deterministic rule skeleton (anchors, decision reversals, mentions, deadlines) + deepseek-v4-flash natural-language fill. Original data: synthetic Microsoft-Teams-style enterprise conversations generated by the GroupMemBench pipeline (GPT-5). Cite the original paper when using the questions. Coverage exceptions (evidence_not_observed violations) are listed in coverage_exceptions.json; low-confidence anchors are marked in questions_enhanced.jsonl.

How to Run

Prerequisites

Set the required environment variable:

export DEEPSEEK_API_KEY="your-deepseek-api-key"

Download Source Data and Run Synthesis

mkdir -p /tmp/gmb && cd /tmp/gmb
curl -sLO "https://raw.githubusercontent.com/UCSB-NLP-Chang/GroupMemBench/main/data/final/Finance/synthetic_domain_channels_rolevariants_Finance.json"
mkdir -p questions && cd questions
for f in multi_hop knowledge_update temporal term_ambiguity user_implicit abstention; do
  curl -sLO "https://raw.githubusercontent.com/UCSB-NLP-Chang/GroupMemBench/main/questions/Finance/$f.jsonl"
done
cd <repo-root>
go run ./cmd/groupmembench-sessions \
  -domain-file /tmp/gmb/synthetic_domain_channels_rolevariants_Finance.json \
  -questions-dir /tmp/gmb/questions \
  -out-dir out/finance -cache-dir out/finance-cache

Expected output: 3389 sessions across 12 users, 0 fallback actions, 134 coverage exceptions (131 lurker-visibility from asker never posting in evidence channel; 3 per-session action-cap truncation), 30-session QA sample passed, team-level coverage complete (every evidence message observed and memory-written by ≥1 member).

Manual Review Before Hugging Face Upload

Before uploading to Hugging Face:

1. Coverage exceptions: Inspect out/finance/coverage_exceptions.json. For each evidence_not_observed entry, determine whether the issue is a real data problem (evidence genuinely not observed by the agent) or an anchor classification error. Real data problems should be documented in this README; anchor errors warrant re-verification with corrected questions_enhanced.jsonl.

2. Low-confidence questions: Manually review all questions marked with "confidence": "low" in questions_enhanced.jsonl. Verify evidence quality and anchor correctness before publishing.

3. Session spot check: Randomly sample 30 sessions and inspect the trajectory quality (persona voice authenticity, factual fidelity to original messages).