lysos-amr-stage2-pro-v2

A 364,432-train + 29,300-valid instruction-tuning dataset for fine-tuning Gemma 4 31B-it on antimicrobial drug design and clinical reasoning, with an elite chain-of-thought (CoT) reasoning slice covering 14 named-drug task types — plus 55 held-out reasoning prompts for independent evaluation.

This is the default Stage 2 SFT corpus for the Lysos project (AMD Developer Hackathon 2026). It supersedes lysos-amr-stage2-pro (v1) by adding a high-quality, named-genes/MIC/trial-cited reasoning slice that targets the under-specified slots in the original dataset.

What's new in v2

Slice	v1	v2	Δ
Total train	364,113	364,432	+319
Total valid	29,300	29,300	0
Held-out test (named-drug eval)	0	55	new
drug_pathogen_reasoning	252	280	+28
drug_mechanism_deep_dive	40	79	+39
counterfactual_design	20	51	+31
resistance_mechanism_explanation	12	46	+34
drug_combination_synergy	17	44	+27
cross_pathogen_spectrum	12	36	+24
structure_activity_comparison	15	36	+21
pathogen_specific_dive	6	32	+26
pk_pd_reasoning	8	32	+24
stewardship_decision	7	30	+23
design_challenge	6	11	+5
candidate_ranking	0	17	new
reward_profile_analysis	0	17	new
structural_strategy_reasoning	0	17	new

14 active reasoning task types plus the original ~13 chemistry/vocabulary task types from v1, totaling ~27 task types in the final mix.

The 3 brand-new task types (candidate_ranking, reward_profile_analysis, structural_strategy_reasoning) plug the Stage 3 GRPO reward-objective gap — they teach the model to reason about multi-objective trade-offs across the 7 reward dimensions (predicted MIC, QED, SA, hemolysis, Tanimoto novelty, embedding novelty, validity).

Schema

Each row has:

{
  "task":     "drug_pathogen_reasoning" | ...,
  "split":    "train" | "valid",
  "prompt":   "<instruction + named context>",
  "response": "<gold reasoning + decision>",
  "messages": "[{\"role\": \"user\", \"content\": ...}, ...]"
}

Note messages is a JSON-encoded string (matches v1 schema for TRL SFTTrainer compatibility with text_field=messages).

Recommended task_mix

The Stage 2 trainer config (configs/stage2_amr_sft.yaml in the Lysos repo) uses these weights, summing to 1.0 across 26 task types:

# Core chemistry (46%)
generation_for_target: 0.18
activity_prediction: 0.16
peptide_design: 0.08
drug_smiles: 0.02
drug_from_smiles: 0.02

# Safety + properties (12%)
safety_prediction: 0.08
drug_likeness: 0.04

# Drug + natural product knowledge (37%)
natural_product_origin: 0.04
natural_product_smiles: 0.03
natural_product_origin_smiles: 0.03
drug_id_lookup: 0.05
drug_inchi_key: 0.05
drug_synonyms: 0.05
drug_cas_lookup: 0.06
drug_reverse_cas: 0.06

# Elite named-drug reasoning slice (5%) — 25x oversample
drug_pathogen_reasoning: 0.020
drug_mechanism_deep_dive: 0.005
counterfactual_design: 0.004
resistance_mechanism_explanation: 0.003
drug_combination_synergy: 0.003
cross_pathogen_spectrum: 0.003
structure_activity_comparison: 0.002
pathogen_specific_dive: 0.002
pk_pd_reasoning: 0.002
stewardship_decision: 0.002
design_challenge: 0.001
candidate_ranking: 0.001
reward_profile_analysis: 0.001
structural_strategy_reasoning: 0.001

The reasoning slice represents 0.2% of the corpus by row count but 5% of training compute — a deliberate 25x oversample to give the model an elite reasoning fingerprint without dominating the chemistry/vocabulary slots.

Held-out test set

49 prompts are held out for independent named-drug-reasoning evaluation. Available as data/synthetic/named_drug_test_split.jsonl in the Lysos repository (also at data/processed/amr-stage2-pro-v2/test_named_drug.jsonl inside the dataset directory).

Stratified by task type (deterministic SHA-256 hash of prompt → top 14% per task), so every reasoning task type has ≥1 held-out example.

The merge script (scripts/merge_named_drug_into_stage2.py) verifies zero leakage: any pre-existing prompt in v1 train that matches a held-out test prompt is filtered out before re-attaching.

Quality controls

The reasoning slice has been QC'd via scripts/qc_named_drug.py:

• ✓ Schema integrity across 388 entries
• ✓ Zero internal duplicate prompts or (prompt, response) pairs
• ✓ Zero held-out test leakage in train
• ✓ Cleaned 2 corrupt entries (truncated/placeholder) detected during QC
• Drug-name resolution: 73 combination drugs (e.g. ceftaz-avi, mero-vab, sul-dur, aztreonam-avibactam) are flagged as "unknown" by the heuristic but are all real combination antibiotics that aren't in the canonical known-antibiotics.parquet index; these are expected false-positives in the QC heuristic.

Provenance

• Reasoning slice authored 2026-04-30 to 2026-05-03 by manual curation against named guidelines (IDSA, ATS, AHA, WHO, CDC), named trials (POET, MERINO, BPaL Nix-TB, ZeNix, ATTACK, Hepburn, van der Horst, Eagle, Geriak, etc.), and named drugs (each entry references DrugBank-resolvable molecules with specific MIC values, named genes/enzymes, and PK/PD targets).
• See data/cot/sprint*.yaml in the Lysos repo for source manifests.
• Schema-converted to Stage 2 pro format via scripts/merge_named_drug_into_stage2.py.

Use

This dataset is intended for the Stage 2 SFT phase of the Lysos training pipeline:

1. Stage 1 — TxGemma-4 (chemistry foundation on TDC tasks)
2. Stage 2 (this dataset) — AMR specialization on top of TxGemma-4
3. Stage 3 — GRPO RL with 7-component reward stack

Loading example:

from datasets import load_dataset
ds = load_dataset("rahul24raj/lysos-amr-stage2-pro-v2")
print(ds["train"][0])

License

Apache 2.0 — same as v1.

Citation

@dataset{lysos_amr_stage2_pro_v2_2026,
  author = {Rahul Rajpurohit},
  title  = {lysos-amr-stage2-pro-v2: AMR drug-design + elite reasoning corpus},
  year   = {2026},
  url    = {https://huggingface.co/datasets/rahul24raj/lysos-amr-stage2-pro-v2},
  note   = {AMD Developer Hackathon 2026, Track 2}
}