AMP Challenge 2027 submission

Local inference artifacts

Place the trained files at these repository paths:

• artifacts/ctmc/model.pt and artifacts/ctmc/config.json: version-2 CTMC checkpoint and its training configuration.
• artifacts/property/model.pt, artifacts/property/model_metadata.json, artifacts/property/backbone/, and artifacts/property/tokenizer/: trained multitask property model, local ESM backbone, tokenizer, and strain metadata.
• data/training/training.fasta and data/training/dataset_manifest.json: disclosed training sequences and their provenance manifest.
• data/antibacterial.fasta: organizer reference sequences.

Generate artifacts/manifest.json after placing all files:

uv run python scripts/build_artifact_manifest.py

The manifest pins every artifact by SHA-256, records the CTMC vocabulary/token IDs and architecture, identifies the training-data manifest, records the official identity source commit, and fixes the seed and output paths. Generation fails before model loading if a required artifact is missing or its hash differs.

Generate

From the repository root, with no required arguments:

uv run generate

The deterministic outputs are generate/library.fasta (50,000 sequences) and generate/top.fasta (100 sequences). Generation loads only local checkpoint/tokenizer files and sets Hugging Face offline mode. It does not train or fetch datasets or model weights.

Validate and rehearse

uv run python scripts/validate_submission.py
uv run python scripts/rehearse_submission.py

The rehearsal runs the normal generate entry point twice in fresh output directories, validates both FASTA pairs, and compares SHA-256 hashes.

Hardware estimate

CUDA is used when available; otherwise inference uses CPU. An RTX 4090-class GPU with 24 GiB VRAM and 32 GiB host memory are planning estimates only; peak memory and generation time have not been measured. CPU generation time has not been measured.

Disclosures

Training-data sources and provenance belong in DATA.md and the referenced data/training/dataset_manifest.json. External model details and limitations belong in MODEL_CARD.md. The implemented method and official-validator provenance are documented in METHOD.md and OFFICIAL_VALIDATOR_SOURCE.md.

Trained CTMC and ESM-2 150M property checkpoints are included under artifacts/ and data/pretrained/esm2_t30_150M_UR50D/ via Git LFS. After a clean clone:

git lfs install
git lfs pull
uv sync
uv run generate
uv run python scripts/validate_submission.py

artifacts/manifest.example.json still contains placeholder hashes and is not a runnable model manifest; use artifacts/manifest.json.

Training exports

Prepare the property-training Parquet table with train/validation/test cluster assignments, then set the exact challenge panel IDs and group membership in configs/train_property.json. Empty panel lists are deliberate placeholders and property export rejects them. The property run exports its checkpoint and local model files to artifacts/property/, and writes the training FASTA and provenance manifest under data/training/.

uv run python scripts/train_property_model.py --data data/processed/observations.parquet --output artifacts/property
uv run python scripts/build_measured_conditions.py --observations data/processed/observations.parquet --split-manifest data/training/split_manifest.csv --property-metadata artifacts/property/model_metadata.json --output artifacts/conditions/measured_conditions.jsonl
uv run python scripts/train_ctmc.py --config configs/train_ctmc.json
uv run python scripts/build_artifact_manifest.py
uv run generate

The default CTMC command uses training_mode: unconditional. Property conditioning uses a separate local prediction cache; property inference is never part of a CTMC epoch:

uv run python scripts/cache_property_conditions.py --checkpoint artifacts/property --fasta data/training/training.fasta --output artifacts/conditions/ctmc_conditions.jsonl --batch-size 32
uv run python scripts/train_ctmc.py --config configs/train_ctmc_property.json

The measured-condition builder uses train-split rows only, aggregates replicates by median, retains assay/censoring/uncertainty audits, and masks conditions that fail panel or group coverage thresholds. Add --measured artifacts/conditions/measured_conditions.jsonl to the cache command to give those measured fields precedence; teacher predictions fill only missing fields. Each wide row has sequence, explicit condition_names in CONDITION_NAMES order, plus values, observed, provenance, and uncertainty maps. Provenance values are measured, teacher, derived, or missing. Activity threshold, temperature, strain groups, and conservative broad aggregation come from property-model metadata. CTMC normalization is fitted only on observed train-split cache entries, stored in the checkpoint, and reused for validation and generation. The cache is create-once and its adjacent .manifest.json records schema, semantics, model hashes, and table SHA-256. Property-conditioned training verifies complete sequence coverage and at least one observation. The CTMC trainer exports to the established artifacts/ctmc/ path. The manifest builder requires the local organizer reference at data/antibacterial.fasta; it never downloads model or data files.