Yangtze-ailab/LDM-CoT-SFT-Qwen3.5-9B-MixedScience

LDM-CoT-SFT-Qwen3.5-9B-MixedScience

A Qwen/Qwen3.5-9B proposer for Large Discovery Models (LDM), trained as the GP-hidden (value-to-language) variant: the surrogate's numerical acquisition values are withheld from the prompt, so the model must infer the epistemic state — which regions are exhausted, which remain uncertain or informative, and whether to exploit or explore — from the observed experience alone.

Model Summary

An LDM runs a recurrent generate → select → evaluate → update loop in which an LLM proposes candidates, a probabilistic surrogate turns observations into a posterior mean and uncertainty, and an acquisition function selects the next experiment. This model is the proposer. It differs from the companion LDM-SFT-Qwen3.5-9B-MixedScience in one respect: the Gaussian-process values are not shown in the prompt during training, which forces the acquisition reasoning to be reconstructed from the raw history in natural language rather than read off surrogate numbers. It emits a chain-of-thought trace followed by a structured action.

• —Base model: Qwen/Qwen3.5-9B
• —Chat template: qwen3_5 (chain-of-thought / thinking enabled)
• —Training data: LDM-CoT-SFT-16K (GP values hidden)

Why hide the GP values

Withholding the surrogate values makes training consistent with deployment, where the acquisition loop exposes no GP numbers to the proposer, and encourages a task-agnostic search intuition that lives in the weights rather than in a numeric input. In practice the acquisition reasoning is barely affected — even when GP values are provided, the reasoning overwhelmingly draws on the observed outcomes (stall length, noise level, best-so-far) rather than on the surrogate numbers.

How to Use

from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "Yangtze-ailab/LDM-CoT-SFT-Qwen3.5-9B-MixedScience"
tok = AutoTokenizer.from_pretrained(repo, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    repo, torch_dtype="bfloat16", device_map="auto", trust_remote_code=True)

messages = [
    {"role": "system", "content": SYSTEM_PROMPT},   # proposer role + output contract
    {"role": "user",   "content": SEARCH_STATE},    # evaluated history + constraints (no GP values)
]
inputs = tok.apply_chat_template(
    messages, add_generation_prompt=True, enable_thinking=True, return_tensors="pt"
).to(model.device)
out = model.generate(inputs, max_new_tokens=2048, temperature=0.7)
print(tok.decode(out[0][inputs.shape[1]:], skip_special_tokens=True))

Training

Full-parameter SFT of Qwen/Qwen3.5-9B (DeepSpeed ZeRO-3 with CPU offload, bf16, gradient checkpointing), sequence length 16,384, learning rate 1e-5 with a cosine schedule and 0.03 warmup ratio, 2 epochs; framework LLaMA-Factory. Chat template qwen3_5 with thinking enabled. The reasoning targets are in English.

Related

• —LDM-SFT-Qwen3.5-9B-MixedScience — companion proposer trained with GP values shown in the prompt.
• —Datasets: LDM-CoT-SFT-16K (GP hidden), LDM-CoT-Acq-SFT-16K (GP shown), LDM-TTS-Base-SFT-19K (no reasoning).

Limitations

• —Designed to operate inside the LDM loop; standalone it proposes candidates but does not evaluate them.
• —Behaviour reflects the oracles used during data collection and the covered domains.

License

Released under the MIT license. The base model Qwen/Qwen3.5-9B remains subject to its own license terms.