Yangtze-ailab/LDM-CoT-SFT-Qwen3.5-9B-MixedScience
Yangtze-ailab/LDM-CoT-SFT-Qwen3.5-9B-MixedScience is a 9 billion parameter Qwen3.5-based language model designed as a proposer for Large Discovery Models (LDM). It is specifically fine-tuned for scientific discovery loops, inferring epistemic states and acquisition reasoning from raw historical data without explicit Gaussian-process values. This model emits a chain-of-thought trace followed by a structured action, making it suitable for automated scientific experimentation and candidate generation.
Loading preview...
Model Overview
This model, LDM-CoT-SFT-Qwen3.5-9B-MixedScience, is a 9 billion parameter language model built upon Qwen/Qwen3.5-9B. It functions as a proposer within a Large Discovery Model (LDM) framework, which involves a recurrent generate → select → evaluate → update loop for scientific experimentation.
Key Differentiator
What sets this model apart is its training methodology: it was fine-tuned on the LDM-CoT-SFT-16K dataset where Gaussian-process (GP) values were intentionally withheld from the prompt during training. This forces the model to infer the epistemic state (e.g., uncertainty, informativeness, exploration vs. exploitation) and reconstruct acquisition reasoning solely from observed experimental history in natural language. This approach ensures consistency with deployment scenarios where GP numbers are not exposed to the proposer, fostering a more task-agnostic search intuition embedded directly in the model's weights.
Capabilities & Training
- Base Model:
Qwen/Qwen3.5-9B - Chat Template: Uses
qwen3_5with thinking enabled, allowing for chain-of-thought (CoT) traces before structured actions. - Training: Full-parameter SFT using DeepSpeed ZeRO-3, bf16, and gradient checkpointing, with a sequence length of 16,384 tokens over 2 epochs. The reasoning targets are in English.
Intended Use
This model is specifically designed to operate within the LDM loop, proposing candidates for scientific discovery. It generates a chain-of-thought trace followed by a structured action, making it suitable for automated experimental design where the model needs to reason about the next best experiment based on past observations.