from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

# Load base model
base_model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-30B-Instruct",
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

# Load LoRA adapter
model_llm2 = PeftModel.from_pretrained(base_model, "erictam721/qwen3-30b-llm2-fds-lora")

# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-30B-Instruct")

# Use for parameter extraction
messages = [
    {"role": "system", "content": "You extract parameters and construct FDS formulas."},
    {"role": "user", "content": "Your query with formula context here"}
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer([text], return_tensors="pt").to(model_llm2.device)

outputs = model_llm2.generate(
    **inputs,
    max_new_tokens=1024,
    temperature=0.3,
    do_sample=False
)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)

Training Details

Base Model: Qwen/Qwen2.5-30B-Instruct
Adapter Type: LoRA (Low-Rank Adaptation)
Role: LLM2 in multi-stage FDS pipeline
Task: Extract parameters and construct final formulas with specific values

Fine-tuned using LoRA for efficient adaptation
Trained on FDS formula dataset with parameter extraction examples
Optimized for extracting geographic regions, time periods, currencies, and other formula parameters
Checkpoint: 270

Citation

If you use this model, please cite the base Qwen model.