BillyWang1/qwen3-8b-tau-sft
TEXT GENERATIONConcurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 5, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold
BillyWang1/qwen3-8b-tau-sft is an 8 billion parameter instruction-tuned model based on Qwen/Qwen3-8B, specifically fine-tuned for the tau-bench retail environment. It is designed as a behavior-priming checkpoint for agent reinforcement learning, teaching the model to act as a non-thinking, one-tool-call-per-turn tau-bench agent. This model excels as a sampled tool-use policy, particularly for research in agent RL within the tau-bench framework.
Loading preview...
Overview
This model, BillyWang1/qwen3-8b-tau-sft, is an 8 billion parameter instruction-tuned variant of Qwen/Qwen3-8B. It serves as an epoch-1 checkpoint for initializing Reinforcement Learning (RL) with GRPO, specifically for the tau-bench retail environment.
Key Characteristics
- Non-thinking Agent: Designed to operate as a one-tool-call-per-turn agent, with a patched Qwen3 chat template that structurally prevents 'think-relapse'.
- Tool Call Format: Emits tool calls as inline Hermes text within the assistant content, supporting a single tool call per turn.
- Training Data: Fine-tuned on 595 successful synthetic retail trajectories from
fuvty/tau-bench-synthetic, with reasoning traces stripped and specific task filtering. - Behavior-Priming: Functions as a 'behavior-priming' checkpoint, teaching the model basic agent actions rather than being a finished policy on its own.
- Performance: While a weak greedy decoder (0.043 on retail-test), it demonstrates strong sampled tool-use policy, achieving 0.637 average success on retail-train with pass@16 ≈ 0.97 under sampling (T=1.0).
Intended Use Cases
- Research Artifact: Primarily intended as a research artifact for tau-bench agent RL.
- RL Initialization: Suitable for initializing GRPO or similar RL processes to develop more robust tau-bench policies.
- Sampled Tool-Use: Best utilized with sampling (e.g., T=1.0) for tool-use scenarios, rather than greedy decoding, due to observed greedy decoding pathologies.