amayuelas/Qwen3.5-9B-MatRL-MT-SFT

VISIONPricing:Input $0.431 / Cached $0.0862 / Output $1.12Concurrent Unit Cost:1Model Size:9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 10, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

amayuelas/Qwen3.5-9B-MatRL-MT-SFT is a 9 billion parameter Qwen3.5 model, cold-started for multi-turn tool-use in crystal-structure inverse design. It is specifically fine-tuned to teach the model to "commit" and close episodes within a turn budget, addressing a common failure in base models. This model serves as a foundational checkpoint for learning tool-use mechanics before chemistry is learned via reinforcement learning, with a context length of 32768 tokens.

Loading preview...

Overview

amayuelas/Qwen3.5-9B-MatRL-MT-SFT is a 9 billion parameter model based on Qwen/Qwen3.5-9B, specifically fine-tuned for multi-turn tool-use in crystal-structure inverse design. This model is a "cold-start" checkpoint, meaning its primary purpose is to teach the agent the mechanics of tool interaction, such as proposing, evaluating, refining, and crucially, committing within a turn budget. It addresses the issue where base models often fail to make submit calls, leading to commitment failures rather than chemistry failures.

Key Capabilities

  • Multi-turn Tool-Use: Trained on 1,119 multi-turn tool-use episodes from the amayuelas/matrl-sft-mt dataset.
  • Commitment Learning: Focuses on teaching the model to successfully "commit" and conclude tool-use episodes.
  • Reasoning Channel: Utilizes a native thinking channel (reasoning_content → <think>) that is preserved across tool calls, essential for its intended behavior.
  • Scalability: Represents the 9B parameter scale arm of a larger research effort, allowing for direct comparison with its 4B counterpart.

Training Details

The model was trained for 3 epochs (420 steps) using Qwen/Qwen3.5-9B as the base, with a sequence length of 16,384 tokens. It employed an AdamW optimizer with a learning rate of 1e-5 and bf16 precision. Training was performed on 8×A100-40GB GPUs, utilizing FSDP and Ulysses context parallelism. Notably, four episodes longer than 16,384 tokens were dropped. This checkpoint is intended to be followed by a reinforcement learning (RL) stage for chemistry learning.

Important Considerations

  • This model is designed to be evaluated with enable_thinking=true to leverage its trained reasoning channel.
  • While Qwen3.5 is VL-capable, this specific model is focused on text-only tool-use, and preprocessor_config.json and video_preprocessor_config.json are included for vLLM serving compatibility.

Finetunes
1 models