agents/hypothesis.py

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

device = torch.device("cpu")
model_id = "TheBloke/Mistral-7B-Instruct-v0.1"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id).to(device)

def generate(summaries):
    context = " ".join([p['summary'] for p in summaries])
    prompt = f"Based on these paper summaries:\n{context}\n\nSuggest a new AI research hypothesis:"
    inputs = tokenizer(prompt, return_tensors="pt").to(device)
    outputs = model.generate(**inputs, max_new_tokens=100)
    text = tokenizer.decode(outputs[0], skip_special_tokens=True)
    return text[len(prompt):].strip()