eval_nato_benchmark.py

"""Evaluate fine-tuned model on NATO-specific benchmark.

This script runs as a Hugging Face Job to evaluate the NATO doctrine model
on domain-specific questions and saves results to the Hub.
"""
# /// script
# requires-python = ">=3.11"
# dependencies = [
#     "transformers>=4.40.0",
#     "torch>=2.0.0",
#     "peft>=0.7.0",
#     "datasets>=2.16.0",
#     "huggingface-hub>=0.20.0",
#     "accelerate>=0.20.0",
# ]
# ///

import json
import os
from datetime import datetime
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
from huggingface_hub import HfApi
import torch

def load_model(adapter_path: str, base_model_path: str = "mistralai/Mistral-7B-Instruct-v0.3"):
    """Load the fine-tuned model with adapter."""
    print(f"Loading base model: {base_model_path}")
    base_model = AutoModelForCausalLM.from_pretrained(
        base_model_path,
        device_map="auto",
        torch_dtype=torch.float16
    )

    print(f"Loading LoRA adapter: {adapter_path}")
    model = PeftModel.from_pretrained(base_model, adapter_path)

    tokenizer = AutoTokenizer.from_pretrained(base_model_path)
    tokenizer.pad_token = tokenizer.eos_token

    return model, tokenizer

def generate_response(model, tokenizer, question: str, max_new_tokens: int = 300):
    """Generate a response to a question."""
    messages = [{"role": "user", "content": question}]

    prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

    with torch.no_grad():
        outputs = model.generate(
            **inputs,
            max_new_tokens=max_new_tokens,
            temperature=0.7,
            top_p=0.9,
            do_sample=True
        )

    response = tokenizer.decode(outputs[0], skip_special_tokens=True)
    if "[/INST]" in response:
        response = response.split("[/INST]")[-1].strip()

    return response

def evaluate_question(model, tokenizer, question_data: dict) -> dict:
    """Evaluate a single question."""
    question = question_data["question"]
    key_concepts = question_data["key_concepts"]

    print(f"\nEvaluating: {question}")
    response = generate_response(model, tokenizer, question)

    # Check for key concepts
    concepts_found = [
        concept for concept in key_concepts
        if concept.lower() in response.lower()
    ]

    score = (len(concepts_found) / len(key_concepts)) * 10 if key_concepts else 0

    return {
        "id": question_data["id"],
        "ajp_source": question_data["ajp_source"],
        "difficulty": question_data["difficulty"],
        "question": question,
        "response": response,
        "key_concepts": key_concepts,
        "concepts_found": concepts_found,
        "score": round(score, 2)
    }

def main():
    """Run NATO benchmark evaluation."""
    print("=" * 70)
    print("NATO Doctrine Model - Benchmark Evaluation")
    print("=" * 70)

    # Configuration
    adapter_model = "AndreasThinks/mistral-7b-nato-doctrine"
    base_model = "mistralai/Mistral-7B-Instruct-v0.3"

    # Load NATO test questions
    print("\nLoading NATO test questions...")
    api = HfApi()

    # Download test questions from the repo
    questions_file = api.hf_hub_download(
        repo_id="AndreasThinks/nato-llm-scripts",
        filename="nato_test_questions.json",
        repo_type="model"
    )

    with open(questions_file, 'r') as f:
        test_questions = json.load(f)

    print(f"Loaded {len(test_questions)} test questions")

    # Load model
    model, tokenizer = load_model(adapter_model, base_model)

    # Run evaluation
    print("\n" + "=" * 70)
    print("Running Evaluation")
    print("=" * 70)

    results = []
    for question_data in test_questions:
        result = evaluate_question(model, tokenizer, question_data)
        results.append(result)
        print(f"  Question {result['id']}: Score = {result['score']}/10")

    # Calculate summary statistics
    total_score = sum(r['score'] for r in results)
    max_score = len(results) * 10
    percentage = (total_score / max_score) * 100

    # Group by difficulty
    basic_results = [r for r in results if r['difficulty'] == 'basic']
    intermediate_results = [r for r in results if r['difficulty'] == 'intermediate']

    basic_score = sum(r['score'] for r in basic_results) / len(basic_results) if basic_results else 0
    intermediate_score = sum(r['score'] for r in intermediate_results) / len(intermediate_results) if intermediate_results else 0

    # Create summary
    summary = {
        "model": adapter_model,
        "base_model": base_model,
        "evaluation_date": datetime.now().isoformat(),
        "total_questions": len(results),
        "total_score": round(total_score, 2),
        "max_score": max_score,
        "percentage": round(percentage, 2),
        "average_score": round(total_score / len(results), 2),
        "basic_avg_score": round(basic_score, 2),
        "intermediate_avg_score": round(intermediate_score, 2),
        "results": results
    }

    # Print summary
    print("\n" + "=" * 70)
    print("EVALUATION SUMMARY")
    print("=" * 70)
    print(f"Total Score: {total_score:.2f}/{max_score} ({percentage:.1f}%)")
    print(f"Average Score: {summary['average_score']:.2f}/10")
    print(f"Basic Questions: {basic_score:.2f}/10")
    print(f"Intermediate Questions: {intermediate_score:.2f}/10")
    print("=" * 70)

    # Save results
    output_file = "nato_benchmark_results.json"
    with open(output_file, 'w') as f:
        json.dump(summary, f, indent=2)

    print(f"\nResults saved to: {output_file}")

    # Upload results to Hub
    token = os.environ.get("HF_TOKEN")
    if token:
        print("\nUploading results to Hub...")
        try:
            api = HfApi(token=token)
            api.upload_file(
                path_or_fileobj=output_file,
                path_in_repo=f"results/nato_benchmark_{datetime.now().strftime('%Y%m%d_%H%M%S')}.json",
                repo_id=adapter_model,
                repo_type="model"
            )
            print("✅ Results uploaded to model repository")
        except Exception as e:
            print(f"⚠️ Could not upload results: {e}")

    print("\n✅ NATO benchmark evaluation complete!")
    return summary

if __name__ == "__main__":
    main()