LDM-CoT-SFT-16K

Supervised fine-tuning (SFT) corpus for Large Discovery Models (LDM): a dataset that distils an acquisition-guided, test-time search policy into a language-model proposer so that a single forward pass emulates a full model-based optimization loop.

Dataset Summary

An LDM couples three components in a recurrent generate → select → evaluate → update loop: an LLM that proposes candidate experiments, a probabilistic surrogate that maps observations to a posterior mean and uncertainty, and an acquisition function that selects the next experiment under that uncertainty. Running this loop at high budget yields trajectories of proposal decisions; each decision, together with the reasoning that justifies it, is rendered as one training example. Fine-tuning on this corpus compiles the expensive search policy into the proposer's weights.

This release provides the chain-of-thought variant in which the surrogate's numerical outputs are withheld from the prompt. The model receives only the observed history and must therefore infer the epistemic state — which regions are exhausted, which remain uncertain or informative, and whether to exploit or explore — from experience rather than by reading acquisition values. We refer to this objective as value-to-language distillation.

Supported Tasks

Examples are pooled from three discovery domains under a shared action schema:

Domain	Search space	Objective
AutoResearch (nanoGPT)	training-code and hyperparameter edits	validation bits-per-byte
Small molecule	SMILES candidates (KRAS)	Vina docking score and predicted activity
Antibody	CDRH3 amino-acid sequences	Absolut binding energy

Dataset Structure

Each record follows the Alpaca schema:

{
  "instruction": "round context: evaluated history, constraints, and task specification",
  "input": "",
  "output": "<think> research-progress reasoning </think>{ JSON action }",
  "system": "system prompt defining the proposer's role and output contract"
}

• instruction — the search state presented to the proposer at one round.
• output — a chain-of-thought trace followed by the emitted action (proposed candidates).
• A single training split is provided (train.jsonl).

Data Collection and Processing

High-budget LDM rollouts and the accompanying reasoning traces were produced by a self-hosted DeepSeek teacher operating inside the LDM framework, then filtered by the empirical acquisition-tilted policy and rendered into the Alpaca schema. Reasoning traces were translated into English.

Related Datasets

Dataset	Reasoning target	Surrogate values in prompt
LDM-TTS-Base-SFT-19K	direct action, no reasoning	—
LDM-CoT-Acq-SFT-16K	chain-of-thought	shown
LDM-CoT-SFT-16K (this)	chain-of-thought	withheld

Intended Use

Full-parameter SFT of an instruction model as a discovery proposer; the intended base is Qwen/Qwen3.5-9B with the qwen3_5 chat template. The resulting model is deployed inside the LDM acquisition loop, where the surrogate and acquisition function remain external.

Limitations

• Reasoning traces are model-generated and machine-translated; a small fraction of records retain residual non-English text.
• Objective values reflect the specific oracles used during collection (Vina, an activity model, and Absolut) and should not be read as experimental ground truth.

Loading

from datasets import load_dataset
ds = load_dataset("Yangtze-ailab/LDM-CoT-SFT-16K")