app.py

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
Hugging Face Space: LLM Benchmarking App using Gradio
- Upload config.yaml and dataset.jsonl
- Select task
- Run benchmarking across multiple models
- Compute metrics: Exact Match, F1, ROUGE-L, BLEU
- Optional judge scoring
- Display results and allow CSV download
"""

import os
import time
import json
import yaml
import gradio as gr
import pandas as pd
from tqdm import tqdm
from huggingface_hub import login

# ---------------- Authentication ---------------- #
HF_TOKEN = (os.environ.get("HUGGINGFACE_HUB_TOKEN", "") or "").strip()
if HF_TOKEN:
    login(token=HF_TOKEN)
else:
    print("⚠️ WARNING: HF_TOKEN not found. Gated models may fail.")

# ---------------- Optional Metrics ---------------- #
try:
    from rouge_score import rouge_scorer
except ImportError:
    rouge_scorer = None

try:
    import sacrebleu
except ImportError:
    sacrebleu = None

# ---------------- Metrics ---------------- #
def exact_match(pred, ref):
    return float(pred.strip().lower() == ref.strip().lower())

def token_f1(pred, ref):
    pred_tokens = pred.lower().split()
    ref_tokens = ref.lower().split()
    if not pred_tokens and not ref_tokens:
        return 1.0
    if not pred_tokens or not ref_tokens:
        return 0.0
    common = sum(min(pred_tokens.count(t), ref_tokens.count(t)) for t in set(pred_tokens))
    precision = common / len(pred_tokens)
    recall = common / len(ref_tokens)
    return 2 * precision * recall / (precision + recall) if precision + recall else 0.0

def rouge_l(pred, ref):
    if rouge_scorer:
        scorer = rouge_scorer.RougeScorer(["rougeL"], use_stemmer=True)
        return scorer.score(ref, pred)["rougeL"].fmeasure
    return 0.0

def bleu(pred, ref):
    if sacrebleu:
        return sacrebleu.corpus_bleu([pred], [[ref]]).score
    return 0.0

def compute_metrics(task, prediction, reference):
    metrics = {}
    if task in ("qa", "classification"):
        metrics["exact_match"] = exact_match(prediction, reference)
        metrics["f1"] = token_f1(prediction, reference)
    elif task in ("summarization", "translation", "conversation"):
        metrics["rougeL_f"] = rouge_l(prediction, reference)
        metrics["bleu"] = bleu(prediction, reference)
    else:
        metrics["f1"] = token_f1(prediction, reference)
    return metrics

# ---------------- Hugging Face Inference ---------------- #
def hf_generate(model_name, prompt, max_new_tokens=256, temperature=0.2):
    from huggingface_hub import InferenceClient
    client = InferenceClient(model=model_name, token=HF_TOKEN)
    start = time.time()
    try:
        # Detect model type for correct endpoint
        if "flan" in model_name or "t5" in model_name:
            output = client.text2text_generation(prompt, max_new_tokens=max_new_tokens)
        else:
            output = client.text_generation(prompt, max_new_tokens=max_new_tokens, temperature=temperature)
        latency = time.time() - start
        return output.strip(), latency
    except Exception as e:
        return f"ERROR: {str(e)}", time.time() - start

# ---------------- Judge Function ---------------- #
def hf_judge(model_name, prompt, candidate, reference=None, rubric=None, max_new_tokens=256):
    from huggingface_hub import InferenceClient
    client = InferenceClient(model=model_name, token=HF_TOKEN)
    rubric = rubric or (
        "Evaluate the candidate answer. Score 1–5 for:\n"
        "- Relevance\n- Factuality\n- Clarity\nReturn JSON: {\"relevance\": int, \"factuality\": int, \"clarity\": int, \"overall\": float}"
    )
    judge_prompt = f"{rubric}\n\nPrompt:\n{prompt}\nCandidate:\n{candidate}\nReference:\n{reference or 'N/A'}"
    try:
        text = client.text_generation(judge_prompt, max_new_tokens=max_new_tokens, temperature=0.0)
        import re
        m = re.search(r'\{.*\}', text, re.S)
        return json.loads(m.group(0)) if m else {"raw": text}
    except Exception as e:
        return {"error": str(e)}

# ---------------- Benchmark Function ---------------- #
def benchmark(config_text, dataset_text, task, use_judge=False):
    cfg = yaml.safe_load(config_text)
    data = [json.loads(line) for line in dataset_text.splitlines() if line.strip()]
    models = cfg.get("models", [])
    templates = cfg.get("prompt_templates", {})
    template = templates.get(task, "{{text}}")
    judge_cfg = cfg.get("judge", {})

    results = []
    for m in models:
        model_name = m["name"]
        max_tokens = m.get("params", {}).get("max_tokens", 256)
        temperature = m.get("params", {}).get("temperature", 0.2)
        for ex in tqdm(data, desc=model_name):
            variables = {k: ex.get(k, "") for k in ("question", "context", "text", "labels")}
            prompt = template
            for k, v in variables.items():
                prompt = prompt.replace(f"{{{{{k}}}}}", str(v))
            prediction, latency = hf_generate(model_name, prompt, max_new_tokens=max_tokens, temperature=temperature)
            metrics = compute_metrics(task, prediction, ex.get("reference", ""))
            row = {
                "model": model_name,
                "id": ex.get("id", ""),
                "task": task,
                "prompt": prompt,
                "prediction": prediction,
                "reference": ex.get("reference", ""),
                "latency_seconds": latency,
                **metrics
            }
            if use_judge and judge_cfg.get("enabled"):
                scores = hf_judge(judge_cfg.get("model"), prompt, prediction, ex.get("reference", ""), judge_cfg.get("rubric"))
                for k, v in (scores.items() if isinstance(scores, dict) else []):
                    row[f"judge_{k}"] = v
            results.append(row)

    df = pd.DataFrame(results)
    summary = []
    for model_name in set(df["model"]):
        sub = df[df["model"] == model_name]
        summary.append(f"## {model_name}")
        summary.append(f"Samples: {len(sub)}")
        for metric in ["exact_match", "f1", "rougeL_f", "bleu", "judge_overall"]:
            if metric in sub.columns:
                vals = [v for v in sub[metric] if isinstance(v, (int, float))]
                if vals:
                    summary.append(f"{metric}: mean={sum(vals)/len(vals):.4f}")
        summary.append(f"Latency mean: {sum(sub['latency_seconds'])/len(sub):.3f}s\n")
    return df, "\n".join(summary)

# ---------------- Gradio UI ---------------- #
with gr.Blocks() as demo:
    gr.Markdown("# LLM Benchmarking App (Hugging Face)")
    gr.Markdown("Upload config.yaml and dataset.jsonl, select task, and run benchmark.")

    with gr.Row():
        config_file = gr.File(label="Upload Config (YAML)", type="filepath")
        dataset_file = gr.File(label="Upload Dataset (JSONL)", type="filepath")

    task = gr.Dropdown(choices=["qa", "summarization", "classification", "conversation"], label="Select Task")
    use_judge = gr.Checkbox(label="Enable Judge Scoring?", value=False)
    run_btn = gr.Button("Run Benchmark")

    results_table = gr.Dataframe(headers=[
        "model","id","task","prompt","prediction","reference","latency_seconds",
        "exact_match","f1","rougeL_f","bleu","judge_overall"
    ], label="Results")

    summary_box = gr.Textbox(label="Summary", lines=10)
    download_csv = gr.File(label="Download CSV")

    def run_benchmark(config_path, dataset_path, task, use_judge):
        if not config_path or not dataset_path:
            return None, "Error: Please upload both files", None
        config_text = open(config_path, "r", encoding="utf-8").read()
        dataset_text = open(dataset_path, "r", encoding="utf-8").read()
        df, summary = benchmark(config_text, dataset_text, task, use_judge)
        csv_path = "results.csv"
        df.to_csv(csv_path, index=False)
        return df, summary, csv_path

    run_btn.click(run_benchmark, inputs=[config_file, dataset_file, task, use_judge], outputs=[results_table, summary_box, download_csv])

demo.launch()