README.md

metadata

license: apache-2.0
tags:
  - tau2-bench
  - trajectories
  - tool-use
  - agents
pretty_name: Florin tau2-retail official-run trajectories

τ²-bench retail official-run trajectories — Florin-tau2-Retail-30B

DRAFT — private staging copy. Do not make public without explicit approval.

Complete, unmodified trajectory file (results.json, tau2-bench native format) for the official τ²-bench retail evaluation of florin-inc/florin-tau2-retail-30b, published for reproducibility in the same way the RAFT-30B-A3B submission shipped its trajectories.

• 114 retail tasks × 4 trials = 456 simulations, seed 300, max_steps 200
• Harness: sierra-research/tau2-bench @ a2c024725189473d2d7cea3a5cfdbcc67478e41f (v1.0.1 task set)
• Agent: base Qwen3-30B-A3B-Thinking-2507 + LoRA adapter, served by vLLM (temperature 1.0, top_p 0.95, max_tokens 8192)
• User simulator: gpt-5.2, reasoning_effort=low (leaderboard-documented protocol)
• Grading: DB-state check × NL-assertion judge (gpt-4.1, hardcoded by the harness)
• Results (tau2 compute_metrics): pass^1 82.46, pass^2 71.93, pass^3 64.91, pass^4 59.65; 0 infrastructure errors; all 456 terminations user_stop
• sha256(results.json) = 1120aa1272848c30ad8c04d939f11d1b66f10a76db673d582dc2ec3ddc14d806

Verify the metrics locally:

pip install -e git+https://github.com/sierra-research/tau2-bench@a2c0247#egg=tau2
python -c "
from tau2.data_model.simulation import Results
from tau2.metrics.agent_metrics import compute_metrics
print(compute_metrics(Results.load('results.json')))
"