Inference

Load the cap-50 checkpoint from the Hugging Face Hub: THGLab/Llama-3.2-1B-Instruct-LinkLlama-Cap50 (same id you pass to AutoModelForCausalLM.from_pretrained). You can instead set sampling.model_path to a local directory in Hugging Face layout if you exported weights offline.

Config YAML

Use data/inference_config.yaml at the repository root: tunable fields under prompt (linker type, rotatable bonds / heavy atoms, H-bond counts, MW / logP / TPSA bands, reasonability) and sampling (model_path, num_samples, max_length). Use null where unconditional.

sampling.temperature and sampling.top_p default to 1.4 and 0.99 (documented in the YAML). They are not overridable from the CLI; change them in your config copy if you need different generation behavior.

Copy and edit locally:

cp data/inference_config.yaml my_run.yaml

Use huggingface-cli login or HF_TOKEN for gated Hub / Llama access when needed.

Run

From the repository root (where pyproject.toml lives):

python linkllama/llm/inference.py --config my_run.yaml --csv your_input.csv

Optional: --num_samples (overrides sampling.num_samples in the YAML).

Input: CSV or SDF per inference.py --help. GPU recommended.

Minimal example (two rows, same fragment convention as benchmark ZINC CSVs): data/zinc_minimal.csv. Quick smoke test (fewer samples than the default in the YAML):

python linkllama/llm/inference.py \
  --config data/inference_config.yaml \
  --csv data/zinc_minimal.csv \
  --num_samples 2

See data/README.md for what lives under data/.

More module detail: linkllama/llm/README.md.