ZuoHaotong/Qwen2.5-3B-Instruct-SFT-in-WebShop

TEXT GENERATIONPricing:Input $0.32 / Cached $0.064 / Output $1.6Concurrent Unit Cost:1Model Size:3.1BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Aug 20, 2026License:otherArchitecture:Transformer0.0K Featherless Exclusive Cold

ZuoHaotong/Qwen2.5-3B-Instruct-SFT-in-WebShop is a 3.1 billion parameter instruction-tuned model, derived from Qwen/Qwen2.5-3B-Instruct, specifically fine-tuned for multi-turn interaction within the WebShop environment. This model is optimized to act as a shopping agent, demonstrating improved success rates and valid-action rates in simulated e-commerce tasks. It excels at following shopping instructions and executing actions like searching and clicking within a structured web environment.

Loading preview...

Overview

This model, ZuoHaotong/Qwen2.5-3B-Instruct-SFT-in-WebShop, is a full-parameter supervised fine-tune (SFT) of the Qwen2.5-3B-Instruct base model. It is specifically designed for multi-turn interactions within the WebShop environment, aiming to function as an automated shopping agent. The released weights correspond to checkpoint global_step_160, which was selected for achieving the highest success rate among evaluated SFT checkpoints on a fixed set of 256 held-out WebShop goals.

Key Capabilities and Performance

  • WebShop Agent: Optimized for navigating and interacting with the WebShop environment to fulfill shopping instructions.
  • Improved Success Rate: Achieved an absolute gain of 3.906 percentage points in success rate on the evaluation set, increasing from 1.56% (baseline) to 5.47%.
  • Enhanced Action Validity: Demonstrated a 12.365 percentage point increase in valid-action rate compared to the baseline.
  • Multi-turn Interaction: Trained on 3,000 environment-verified teacher trajectories with an average of 5.164 interaction turns per trajectory.
  • Specific Interaction Format: Expects a system instruction defining it as a WebShop shopping agent and assistant turns structured with <think>rationale</think><answer>action</answer>.

Training Details

The model was trained for 1 epoch using a learning rate of 5e-6 and BF16 precision, with a maximum sequence length of 8,192. The training data comprised 1,505 unique training rows and 167 validation rows, ensuring an ASIN-disjoint split. The teacher data generation involved qwen3.6-35b-a3b and webshop-small, with strict quality gates for trajectory acceptance.

Limitations

  • The absolute held-out success rate is 5.47%, indicating it's an initialization for further Agent RL research rather than a production system.
  • Performance metrics are point estimates from a single evaluation set, without statistical significance claims.
  • Specialized to the WebShop environment's formatting and action space, and not intended for generalization to real-world e-commerce sites.