ahammad115566/smeft-qwen-14b

Example:

prompt = """
What are the independent CP-even dimension-6 operators in the Warsaw basis that
modify the triple gauge couplings?
"""

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_use_double_quant=True,
)

model = AutoModelForCausalLM.from_pretrained(
    "ahammad115566/smeft-qwen-14b",
    quantization_config=bnb_config,
    device_map="auto",
    trust_remote_code=True,
)
tokenizer = AutoTokenizer.from_pretrained(
    "ahammad115566/smeft-qwen-14b", 
    trust_remote_code=True
)

def ask(instruction: str, context: str = "") -> str:
    parts = [f"### Instruction:\n{instruction}"]
    if context.strip():
        parts.append(f"\n### Input:\n{context}")
    parts.append("\n### Response:\n")
    prompt = "\n".join(parts)

    inputs = tokenizer(prompt, return_tensors="pt").to(DEVICE)

    with torch.no_grad():
        output_ids = model.generate(
            **inputs,
            max_new_tokens=1024,
            temperature=0.1,
            do_sample=True,
            pad_token_id=tokenizer.eos_token_id,
            repetition_penalty=1.3,       # penalises repeating prior tokens 
            no_repeat_ngram_size=8,   
        )

    new_tokens = output_ids[0][inputs["input_ids"].shape[-1]:]
    return tokenizer.decode(new_tokens, skip_special_tokens=True).strip()

print(ask(prompt))

Training Details

Parameter	Value
Base model	Qwen2.5-14B-Instruct
Fine-tuning method	LoRA (merged)
Quantization at inference	4-bit NF4 (bitsandbytes)
Domain	Standard Model Effective Field Theory
Training data source	SMEFT particle physics preprints

Citation

A paper describing this model and dataset is forthcoming. In the meantime, please cite:

@misc{hammad2025smeft,
  author       = {Ahmed Hammad and Veronica Sanz},
  title        =  To be announced
  year         = {2026},
  publisher    = ,
  url          = {https://huggingface.co/ahammad115566/smeft-qwen-14b}
}

Intended Use and Limitations

Intended use: Research assistance in SMEFT and related EFT frameworks. This model is designed to support physicists working on BSM phenomenology, operator basis manipulation, and Wilson coefficient analysis.

Limitations: The model may occasionally produce plausible-sounding but incorrect numerical values or operator identifications. All results should be verified against the primary literature. The model is not a substitute for rigorous calculation.

License

This model is released under the Apache 2.0 License. The base model (Qwen2.5-14B-Instruct) is subject to its own license terms. """