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

Name: LDM-Acq-SFT-Qwen3.5-9B-MixedScience
Rating: 2.5 (8 reviews)

A Qwen/Qwen3.5-9B model fine-tuned to serve as the proposer inside a Large Discovery Model (LDM): given the state of an ongoing optimization campaign, it reasons about the search progress and proposes the next candidate experiment, amortising a high-budget model-based search loop into a single forward pass.

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, trained by full-parameter supervised fine-tuning on trajectories collected from that loop across three scientific-discovery domains, so that the acquisition-guided search policy is distilled into its weights. It emits a chain-of-thought trace followed by a structured action (the proposed candidates).

Base model: Qwen/Qwen3.5-9B
Chat template: qwen3_5 (chain-of-thought / thinking enabled)
Domains: AutoResearch (nanoGPT), small-molecule design, antibody (CDRH3) design

Intended Use

Deployment as the candidate proposer within the LDM acquisition loop, where the surrogate and acquisition function remain external. The model reads the evaluated history and task constraints and returns reasoning plus the next candidate(s).

How to Use

python

from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "Yangtze-ailab/LDM-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
]
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))

The output is a <think> … </think> reasoning block followed by a JSON action describing the proposed candidate(s).

Training

Method: full-parameter SFT (DeepSpeed ZeRO-3 with CPU offload, bf16, gradient checkpointing)
Sequence length: 16,384
Optimisation: learning rate 1e-5, cosine schedule, warmup ratio 0.03, 2 epochs, effective batch size = per-device 1 × gradient accumulation 8 × #GPUs
Framework: LLaMA-Factory

Training Data

Fine-tuned on the LDM mixed-science SFT corpus — proposal decisions collected from high-budget LDM test-time search across the three domains and rendered in Alpaca format. The corresponding public datasets are:

LDM-CoT-Acq-SFT-16K — chain-of-thought, surrogate values shown in the prompt
LDM-CoT-SFT-16K — chain-of-thought, surrogate values hidden
LDM-TTS-Base-SFT-19K — direct action, no reasoning

Limitations

The model is designed to operate inside the LDM loop; used standalone it proposes candidates but does not itself evaluate them.
Behaviour reflects the specific oracles used during data collection (Vina, an activity model, and Absolut) and the three covered domains; transfer beyond them is not guaranteed.
Reasoning traces in the training data were machine-generated and translated to English.

License

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

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