---
base_model: meta-llama/Llama-3.2-1B-Instruct
license: llama3.2
tags:
  - knowledge-graph-completion
  - link-prediction
  - biomedical
  - cold-start
  - llm-as-scorer
pipeline_tag: text-generation
---

# Cold-Start Chemical–Gene Ranking Readout (1B)

**Built with Llama.** Full fine-tune of **`meta-llama/Llama-3.2-1B-Instruct`** from the paper
*"Cold-Start Link Prediction Needs a Ranking Readout"* (GaLM 2026 @ CIKM). The 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.

This is the **reproducible headline model**: a *generic* (non-biomedical) Llama-3.2-1B,
fine-tuned with a retained recipe, that reaches the same numbers as the biomedical 1B used in
the paper — showing the result rides **general pretraining**, not domain-specific leakage.

## Headline results (cold-start, full candidate set of 13,972 genes)
| Metric | Value |
|---|---|
| Full-candidate MRR | **0.536** |
| Sampled MRR (random neg, K=99) | 0.837 |
| Sampled MRR (hard neg, K=99) | 0.798 |
| Popularity floor (full-candidate) | 0.106 |

The score is `s(h,r,g) = (1/|g|) · log P_θ(g | π(h,r))`; genes are ranked by `s`.
Candidate-constrained decoding is equivalent to this ranking readout; **unconstrained
generation fails** at cold-start, which is the paper's central point.

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

repo = "BioRel/coldstart-ranking-readout-1b"
tok = AutoTokenizer.from_pretrained(repo)
lm  = AutoModelForCausalLM.from_pretrained(repo, torch_dtype=torch.bfloat16, device_map={"": 0}).eval()

# rank genes by length-normalized log-prob of each candidate given the (chemical, relation) question:
# s(q, g) = (1/|g|) * sum_t log P(g_t | q, g_<t)
# Full scorer: https://github.com/<repo>  (src/06_baselines/path_b_llm_ranking.py, kg_llm_exactrank.py)
```

## Intended use & limitations
- **Research use only.** Cold-start chemical–gene *ranking* (scoring), not free generation and
  not clinical decision-making.
- The companion LoRA adapters`BioRel/coldstart-lora-{1b,3b,8b}` is also provided.

## Training data & license
- **Data:** derived from the **Comparative Toxicogenomics Database (CTD)**, https://ctdbase.org.
  **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-3.2-1B-Instruct` — governed by the **Llama 3.2 Community
  License**. You must accept Meta's license. This fine-tune is distributed under the same terms.

Please cite CTD: A. P. Davis et al., *Comparative Toxicogenomics Database (CTD): update 2021*,
Nucleic Acids Research, 2021. https://ctdbase.org