text2graph-llama-3.2-3b

This model is a fine-tuned version of unsloth/Llama-3.2-3B-Instruct on a text2triple dataset curated by Sonnet-3.5.

This model has much faster inference speed than our previous trained T5-based model. Also, it performs better for longer (> 512 tokens) input.

Example Input:

"William Gerald Standridge (November 27, 1953 – April 12, 2014) was an American stock car racing driver. He was a competitor in the NASCAR Winston Cup Series and Busch Series."

Example Output:

(S> William gerald standridge| P> Nationality| O> American),
(S> William gerald standridge| P> Occupation| O> Stock car racing driver),
(S> William gerald standridge| P> Competitor| O> Busch series),
(S> William gerald standridge| P> Competitor| O> Nascar winston cup series),
(S> William gerald standridge| P> Birth date| O> November 27, 1953),
(S> William gerald standridge| P> Death date| O> April 12, 2014)

How to Use?

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "pat-jj/text2graph-llama-3.2-3b"

def load_model_and_tokenizer():
    # Load the model and tokenizer
    model = AutoModelForCausalLM.from_pretrained(
        model_name,
        torch_dtype=torch.bfloat16,
        device_map="auto"
    )
    tokenizer = AutoTokenizer.from_pretrained(model_name)
    
    # Set up chat template
    tokenizer.chat_template = tokenizer.chat_template or "llama-3.1"
    
    return model, tokenizer

def generate_triples(model, tokenizer, input_text, max_length=2048):
    # Format the input using chat template
    messages = [{
        "role": "user",
        "content": f"Convert the following text to triples:\n\nText: {input_text}"
    }]
    
    prompt = tokenizer.apply_chat_template(
        messages,
        tokenize=False,
        add_generation_prompt=True
    )
    
    # Tokenize input
    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
    
    # Generate response
    outputs = model.generate(
        **inputs,
        max_length=max_length,
        num_return_sequences=1,
        temperature=0.7,
        do_sample=True,
        pad_token_id=tokenizer.pad_token_id
    )
    
    # Decode and return the response
    response = tokenizer.decode(outputs[0], skip_special_tokens=True)
    return response

def main():
    print("Loading model and tokenizer...")
    model, tokenizer = load_model_and_tokenizer()
    
    print("\nModel loaded! Enter text to convert to triples (type 'quit' to exit):")
    
    while True:
        user_input = input("\nEnter text: ")
        if user_input.lower() == 'quit':
            break
            
        print("\nGenerating triples...")
        response = generate_triples(model, tokenizer, user_input)
        print("\nResponse:", response)

if __name__ == "__main__":
    main()

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 3
• eval_batch_size: 1
• seed: 42
• distributed_type: multi-GPU
• num_devices: 3
• gradient_accumulation_steps: 3
• total_train_batch_size: 27
• total_eval_batch_size: 3
• optimizer: Use adamw_torch with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 1

Framework versions

• Transformers 4.48.1
• Pytorch 2.1.2+cu121
• Datasets 2.21.0
• Tokenizers 0.21.0

Cite Our Paper

@misc{jiang2025rasretrievalandstructuringknowledgeintensivellm,
      title={RAS: Retrieval-And-Structuring for Knowledge-Intensive LLM Generation}, 
      author={Pengcheng Jiang and Lang Cao and Ruike Zhu and Minhao Jiang and Yunyi Zhang and Jimeng Sun and Jiawei Han},
      year={2025},
      eprint={2502.10996},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2502.10996}, 
}