SeanWang0027/qwen3-1.7b-textcraft-rose-s5t5-qwen3-32b-step66

TEXT GENERATIONPricing:Input $0.32 / Cached $0.064 / Output $1.6Concurrent Unit Cost:1Model Size:2BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Sep 20, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

SeanWang0027/qwen3-1.7b-textcraft-rose-s5t5-qwen3-32b-step66 is a 2 billion parameter Qwen3-based language model, fine-tuned using an online multi-turn ROSE (Reward-Optimized Supervised Experience) method with a Qwen3-32B teacher. It is specifically optimized for multi-turn agent interactions within the TextCraft environment, demonstrating a 54.75% success rate on the TextCraft official test split. This model is designed for tasks requiring sequential decision-making and interaction in text-based environments.

Loading preview...

Overview

SeanWang0027/qwen3-1.7b-textcraft-rose-s5t5-qwen3-32b-step66 is a 2 billion parameter model based on the Qwen3-1.7B architecture. It was developed as part of a study comparing different training methodologies for multi-turn agents, specifically focusing on online multi-turn ROSE (Reward-Optimized Supervised Experience).

Key Training Details

  • Methodology: Online multi-turn ROSE, where the student (Qwen3-1.7B) plays the first 5 turns, and a Qwen3-32B teacher continues for up to 5 turns. Only the teacher's turns are used as cross-entropy targets.
  • Episodes: 32 episodes per step, trained over 66 steps (6 epochs).
  • Teacher Model: Qwen3-32B.
  • Context: 512 tokens per turn.

Performance

  • Evaluation: Tested on the TextCraft official test split (100 tasks) with an average of 4 passes.
  • Success Rate: Achieved 54.75% success, significantly outperforming the base Qwen3-1.7B (23.00%) but below the Qwen3-32B teacher (85.50%).

Prompt Format

  • Utilizes a ReAct, multi-turn chat format.
  • Expects the environment's AgentGym instruction as the first user turn, followed by an assistant acknowledgement, and then one user turn per observation.
  • The model responds with Thought:\n...\n\nAction:\n<one action>.

Use Cases

This model is particularly well-suited for applications requiring:

  • Multi-turn agent interactions in text-based environments.
  • Sequential decision-making tasks.
  • TextCraft-like environments where a smaller model needs to emulate the behavior of a larger teacher model.

Finetunes
1 models