Auto-ChemInstruct is a synthetically generated instruction dataset of DPO/RLHF preference pairs for chemistry language models. Each pair consists of a chosen (physically valid) and a rejected (physically invalid) chemical reaction, accompanied by a causal reflection trace explaining why the reaction failed — enabling chain-of-thought reasoning in downstream DSLM fine-tuning.

The dataset is produced by an autonomous multi-agent pipeline with a self-bootstrapping innovation loop: hypothesis generation → physical verification → causal reflection → accumulated learning → repeat with learned constraints.

Key Statistics

Metric	Value
Preference pairs	21 (train: 16, val: 2, test: 3)
Unique molecules	50
Tanimoto diversity	87.5%
Scaffold diversity	42.0%
Pass rate	69.7%
Reflection traces	100%
Reaction types	1+

Data Format

Each JSONL record is a DPO preference pair:

json

{
  "chosen": "<reaction SMILES | reaction description>",
  "rejected": "<reaction SMILES | reaction description> + causal failure analysis>",
  "reaction_type": "amide_coupling | esterification | ...",
  "quality_score": 0.65,
  "verification": {
    "status": "passed | failed",
    "failure_categories": ["kinetic_barrier", "thermodynamic_instability"]
  },
  "reflection": "Causal reasoning trace explaining failure mechanism..."
}

Source

Generated by the Auto-ChemInstruct pipeline using Fireworks AI (deepseek-v4-pro) as the LLM backbone, with RDKit for structural verification and chemical feasibility filtering.

Pipeline: Hypothesis Agent → Verification Agent → Reflection Agent → Compilation Agent
Self-bootstrapping: Cosine temperature annealing (1.0→0.3) across bootstrap iterations
Verification: RDKit structural validation + chemical feasibility (unstable group detection, ring strain, hypervalent atoms)

Intended Use

This dataset is designed for RLHF/DPO fine-tuning of chemistry-focused language models. The preference pairs with causal reflection traces enable models to learn:

Physical validity of molecular reactions
Causal reasoning about why reactions fail
Domain-specific chemical constraints

Citation

bibtex

@dataset{autochem-instruct,
  author = {Kumar, Aayush},
  title = {Auto-ChemInstruct: Agent-Driven Synthesis of RLHF Data for Chemistry DSLMs},
  year = {2026},
  publisher = {HuggingFace},
  url = {https://huggingface.co/datasets/aayushkrm/autochem-instruct}
}