You can deploy this model and start earning money today!

phiwi/
Meta-Llama-3.1-8B-Instruct-bnb-4bit_regulatome_ppi_lora

Name: Meta-Llama-3.1-8B-Instruct-bnb-4bit_regulatome_ppi_lora
Availability: InStock

transformers

safetensors

text-generation-inference

Views

No views yet

Llama 3.1 8B — RegulaTome PPI (LoRA)

Fine-tuned Llama 3.1 8B Instruct (4-bit QLoRA) for protein–protein interaction extraction from biomedical abstracts.

Model Details

Property	Value
Base model	`unsloth/Meta-Llama-3.1-8B-Instruct-bnb-4bit`
Fine-tuning method	LoRA (r=16, alpha=16, dropout=0)
Training quantization	4-bit (NF4)
Inference quantization	Q4_K_M (4.9 GB GGUF)
Chat template	ChatML (`<
Task	PPI triple extraction (INTERACTS_WITH)

Training Data

Corpus: RegulaTome PPI annotations (CC BY 4.0)
Train: 942 abstracts with ground-truth PPI relations
Devel: 337 abstracts (merged with train, 10% held for per-epoch eval)
Test: 312 abstracts (never seen during training)
Total training samples: 1279

Training Configuration

Parameter	Value
Epochs	5
Per-device batch size	2
Gradient accumulation	4 steps
Effective batch size	8
Learning rate	2e-4 (linear schedule)
Warmup steps	5
Optimizer	AdamW 8-bit
Max sequence length	120,000
Hardware	1x NVIDIA A40 (44 GB), ~7.6 GB VRAM
Training time	~50 minutes

Prompt Format

The model was fine-tuned using the ChatML template. For inference, format your prompt as:

<|im_start|>system
You are an expert molecular biologist specializing in protein-protein interactions. Your TASK is to extract protein-protein interactions from scientific texts with high precision. You understand the difference between direct physical interactions, functional relationships, and regulatory effects. When extracting relationships, focus on evidence-based direct interactions rather than indirect associations.<|im_end|>
<|im_start|>system
TEXT: [your biomedical abstract here]<|im_end|>
<|im_start|>system
Use the following OUTPUT FORMAT:
{
    triples: [
    {
        head: string,
        relation: "INTERACTS_WITH",
        tail: string,
    }
    ],
}<|im_end|>
<|im_start|>user
TASK: Extract all the protein-protein interactions from the TEXT. Focus on direct physical interactions where proteins bind to each other, modify each other, or form complexes.<|im_end|>
<|im_start|>assistant

The model responds with a JSON object containing extracted triples.

Usage

Via Ollama

ollama run llama3.1:8b-regulatome-ppi

Via HuggingFace

python

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="phiwi/Meta-Llama-3.1-8B-Instruct-bnb-4bit_regulatome_ppi_lora",
    max_seq_length=120000,
    load_in_4bit=True,
)

messages = [
    {"role": "system", "content": "You are an expert molecular biologist..."},
    {"role": "system", "content": "TEXT: [your abstract]"},
    {"role": "system", "content": "Use the following OUTPUT FORMAT: {...}"},
    {"role": "user", "content": "TASK: Extract all the protein-protein interactions..."},
]
text = tokenizer.apply_chat_template(messages, tokenize=False)

With LINDA-LLM pipeline

python extract.py --model llama31regu --data regulatome --target ppi

Performance

Evaluated on the RegulaTome test set (312 abstracts) across 15 extraction configurations (direct/nerrel x normal/neg/pos/negpos/dynex/lookup/ensemble/tot). See LLM-PPI-Extraction for full results.

Companion Model

A larger 70B variant is also available: phiwi/Llama-3.3-70B-Instruct-bnb-4bit_regulatome_ppi_lora

Citation

bibtex

@software{LLM-PPI-Extraction,
  author = {Philipp Wiesenbach},
  title = {LLM-PPI-Extraction: LLM-based extraction of protein-protein interactions},
  year = {2026},
  url = {https://github.com/dieterich-lab/LLM-PPI-Extraction}
}