README.md

---
base_model: meta-llama/Llama-3.2-3B-Instruct
library_name: peft
license: llama3.2
tags:
  - lora
  - peft
  - knowledge-graph-completion
  - link-prediction
  - biomedical
  - cold-start
pipeline_tag: text-generation
---

# Cold-Start Chemical–Gene Ranking — LoRA adapter (3B)

**Built with Llama.** This is a LoRA (r=64) adapter for **`meta-llama/Llama-3.2-3B-Instruct`**, from the paper
*"Cold-Start Link Prediction Needs a Ranking Readout"* (GaLM 2026 @ CIKM). The fine-tuned
model is read out as a **length-normalized sequence log-probability scorer** to rank candidate
genes for a chemical–gene interaction query, under a cold-start (unseen-chemical) protocol.

## What this is
- **Adapter only** (~371 MB). The base model `meta-llama/Llama-3.2-3B-Instruct` is **not** included — download it from the
  Hugging Face Hub (subject to the Llama Community License).
- Fine-tuned on a **CTD-derived** chemical–gene QA corpus (augmented *sample-47* footing).
- Part of a size ladder released with the paper — **1B 0.78 / 3B 0.86 / 8B 0.92** (cold-start,
  sampled hard-negative MRR, K=99, sample-47). Under this capacity-limited LoRA regime, the
  readout improves monotonically with size, which — together with the full-fine-tuning result
  where a 1B already reaches the ceiling — locates the operative axis at **capacity, not scale**.

## This adapter
- **Cold-start sampled hard-negative MRR = 0.863** (ep8; sample-47 corpus, K=99).
- LoRA config: r=64, alpha=64, target modules q/k/v/o/gate/up/down_proj.

## Intended use & limitations
- **Research use only.** Cold-start chemical–gene *ranking* (scoring), not free generation and
  not clinical decision-making. Absolute values are only comparable **within** the sample-47,
  LoRA footing (never cross-compared with the full-fine-tuning / gl47 headline numbers).

## Usage (sketch)
```python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
from peft import PeftModel

base_id = "meta-llama/Llama-3.2-3B-Instruct"
tok  = AutoTokenizer.from_pretrained("BioRel/coldstart-lora-3b")
bnb  = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_compute_dtype=torch.bfloat16, bnb_4bit_quant_type="nf4")
base = AutoModelForCausalLM.from_pretrained(base_id, quantization_config=bnb, device_map={"": 0})
lm   = PeftModel.from_pretrained(base, "BioRel/coldstart-lora-3b").eval()

# score a candidate gene g for query q by length-normalized log-prob of g given q
# s(q, g) = (1/|g|) * sum_t log P(g_t | q, g_<t)  ; rank genes by s.
# Full scorer: https://github.com/BioRel/relational-qa-coldstart  (src/06_baselines/lora_scorer.py)
```

## Training data & license
- **Data:** derived from the **Comparative Toxicogenomics Database (CTD)**, https://ctdbase.org.
  The dataset is **subject to CTD terms**; users must download CTD data themselves
  (terms: https://ctdbase.org/about/legal.jsp). **CTD may access this dataset for quality control
  purposes.** Non-commercial / research use.
- **Base model:** `meta-llama/Llama-3.2-3B-Instruct` — governed by the **Llama Community License** (llama3.2). You must accept
  Meta's license to download the base.
- **Adapter weights:** released for research use, subject to the base-model and CTD terms above.

## Citation
```bibtex
@inproceedings{kim2026coldstart,
  title     = {Cold-Start Link Prediction Needs a Ranking Readout},
  author    = {Kim, Yunha and Kim, Young-Hak and Jun, Tae Joon},
  booktitle = {Proceedings of the Workshop on Graph-Augmented LLMs (GaLM), co-located with CIKM},
  year      = {2026}
}
```
Please also cite CTD: A. P. Davis et al., *Comparative Toxicogenomics Database (CTD): update 2021*,
Nucleic Acids Research, 2021. https://ctdbase.org