Retraining

1. Corpus (ChEMBL → JSONL)

With conda env and pip install -e .:

python linkllama/llm/fragmentize.py input.smi --output-csv frags.csv
python linkllama/llm/calculate_properties.py frags.csv props.csv
python linkllama/llm/sft_corpus.py props.csv -o corpus.jsonl

Large files live under data/ (gitignored). The cap-50 training file used in the shipped Axolotl config is:

data/chembl_processed/chembl36_balanced_cap50.jsonl

Build it with linkllama/llm/create_balanced_subset.py then sft_corpus.py — see linkllama/llm/README.md.

If you publish this JSONL as a separate Hugging Face dataset repository, point training configs at the downloaded path or your Hub dataset id after snapshot_download.

2. LoRA (Axolotl)

Separate env with Axolotl, accelerate, and bitsandbytes. Export HF_TOKEN for gated Llama weights; never commit tokens.

cd linkllama/training
accelerate launch -m axolotl.cli.train cf_lora_cap50.yml
python -m axolotl.cli.merge_lora cf_lora_cap50.yml --lora_model_dir ./outputs_cap50

Config: cf_lora_cap50.yml.

3. Inference

After merging (or downloading from Hub), set sampling.model_path in data/inference_config.yaml (or your copy of it) to your Hub model id or local path. See inference_guide.md.