hack1984-lora1

HACK1984 Track 1 (Signal-to-Dose) labeler: Qwen3.5-2B fine-tuned with a LoRA adapter (lora1). It labels public evidence on two lung cancer drugs with signal type, actionable and five aspects, and runs offline on a self-hosted llama.cpp server.

Files

File	What it is
lora1-q8_0.gguf	Base model with the lora1 adapter merged, quantized to Q8_0 (2.0 GB). This is the file the pipeline uses.
lora1-adapter/	The LoRA adapter on its own (PEFT format, with tokenizer), for use on top of unsloth/Qwen3.5-2B.

lora1-q8_0.gguf sha256: a5ae0bc319e3331aa05c087a2632e5cc2ba765b2485cf1ace0a4c6468761c238

Use

Put lora1-q8_0.gguf at drug track/models/lora1-q8_0.gguf in the code repo, check the sha256, and follow drug track/RUN.md there. Only the llama-server process touches the model, and the run needs no network.

Training data

Trained only on public, unlabeled track items. No test items were used.