from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline
import torch

model = AutoModelForCausalLM.from_pretrained(
    "Gautamo1/mistral-7b-rag-reader",
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("Gautamo1/mistral-7b-rag-reader")

messages = [
    {"role": "system", "content": "Answer using ONLY the context provided."},
    {"role": "user",   "content": "Context:\n{chunk}\n\nQuestion: {question}"},
]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
output = model.generate(**inputs, max_new_tokens=250, do_sample=False)
answer = tokenizer.decode(output[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)
print(answer)

Training data

Generated from PDF documents using a multi-step pipeline: PDF parsing → question generation → answer generation → hard negative mining → quality filtering.