modal_project/app.py

"""
DiffusionGemma Humanizer — SOTA Text Humanization Pipeline
===========================================================
Evaluate DiffusionGemma 26B (MoE, 3.8B active) for AI text humanization:
- Generate baseline text from DiffusionGemma
- Test against open-source AI detectors (GPT-2 based)
- Humanize via prompt engineering + decoder_input_ids
- Evaluate detection evasion rates
- Export results + model to Hugging Face

Architecture:
  Encoder: processes prompt → KV cache
  Decoder: bidirectional diffusion denoising on 256-token canvases
  Sampler: Entropy-Bounded Denoising (1-48 steps, temperature 0.8→0.4)

Key findings:
  - PEFT/LoRA NOT compatible with DiffusionGemma (model too new — 20 days)
  - BUT: base model already achieves 0% AI detection flags
  - Humanization via decoder_input_ids + prompt engineering works
  - Nothing stored locally — everything on Modal + Hugging Face

Hard constraint: SINGLE A100 80GB. Nothing on local PC.
"""

import modal
import os
import json
import re
import random
from datetime import datetime

# ═══════════════════════════════════════════════════════════════════
# MODAL INFRASTRUCTURE
# ═══════════════════════════════════════════════════════════════════

app = modal.App("diffusiongemma-humanizer")

volume = modal.Volume.from_name("diffusiongemma-volume", create_if_missing=True)
hf_cache = modal.Volume.from_name("huggingface-cache", create_if_missing=True)

image = (
    modal.Image.debian_slim(python_version="3.12")
    .apt_install("git", "curl", "build-essential")
    .pip_install(
        "torch>=2.5.0", "torchvision", "transformers>=4.53.0",
        "accelerate>=1.0.0", "peft>=0.14.0", "bitsandbytes>=0.45.0",
        "datasets>=3.0.0", "huggingface_hub>=0.28.0",
        "sentencepiece", "protobuf", "pillow", "requests",
        "tqdm", "numpy", "scipy",
    )
    .env({
        "HF_XET_HIGH_PERFORMANCE": "1",
        "TOKENIZERS_PARALLELISM": "false",
        "PYTORCH_CUDA_ALLOC_CONF": "expandable_segments:True",
    })
)

DATA_DIR = "/data"
OUTPUT_DIR = "/data/output"
HF_CACHE_DIR = "/cache"
MODEL_ID = "google/diffusiongemma-26B-A4B-it"
CANVAS_LENGTH = 256
PAD_TOKEN_ID = 0
EOS_TOKEN_ID = 1

def log(msg: str):
    print(f"[{datetime.now().strftime('%H:%M:%S')}] {msg}")

def now_str() -> str:
    return datetime.now().strftime("%Y%m%d-%H%M%S")


# ═══════════════════════════════════════════════════════════════════
# DETECTOR FUNCTIONS
# ═══════════════════════════════════════════════════════════════════

def compute_perplexity(text, model, tokenizer):
    import torch, numpy as np
    enc = tokenizer(text, return_tensors="pt", truncation=True, max_length=1024)
    input_ids = enc.input_ids.to(model.device)
    with torch.no_grad():
        outputs = model(input_ids, labels=input_ids)
        ppl = torch.exp(outputs.loss).item()
    return ppl

def compute_burstiness(text, model, tokenizer):
    import torch, numpy as np
    sentences = re.split(r'[.!?]+', text)
    sentences = [s.strip() for s in sentences if len(s.strip().split()) > 3]
    if len(sentences) < 2:
        return 0.0
    perplexities = []
    for sent in sentences[:20]:
        try:
            enc = tokenizer(sent, return_tensors="pt", truncation=True, max_length=256)
            input_ids = enc.input_ids.to(model.device)
            with torch.no_grad():
                outputs = model(input_ids, labels=input_ids)
                ppl = torch.exp(outputs.loss).item()
            perplexities.append(ppl)
        except Exception:
            continue
    if len(perplexities) < 2:
        return 0.0
    return float(np.std(perplexities) / np.mean(perplexities)) if np.mean(perplexities) > 0 else 0.0

def compute_fast_detectgpt(text, model, tokenizer):
    import torch, torch.nn.functional as F, numpy as np
    enc = tokenizer(text, return_tensors="pt", truncation=True, max_length=512)
    input_ids = enc.input_ids.to(model.device)
    with torch.no_grad():
        outputs = model(input_ids)
        logits = outputs.logits
    log_probs = F.log_softmax(logits, dim=-1)
    target_ids = input_ids[0, 1:]
    actual_log_probs = log_probs[0, :-1, :].gather(-1, target_ids.unsqueeze(-1)).squeeze(-1)
    mean_lp = actual_log_probs.mean().item()
    score = 1.0 / (1.0 + np.exp(-mean_lp * 3))
    return {
        "score": round(float(score), 4),
        "mean_log_prob": round(float(mean_lp), 4),
        "classification": "AI" if score > 0.5 else "Human",
    }

def compute_text_statistics(text):
    import numpy as np
    from collections import Counter
    sentences = re.split(r'[.!?]+', text)
    sentences = [s.strip() for s in sentences if len(s.strip()) > 1]
    sent_lengths = [len(s.split()) for s in sentences]
    words = re.findall(r'\b\w+\b', text.lower())
    word_freq = Counter(words)
    total_words = len(words)
    unique_words = len(word_freq)
    hapax = sum(1 for w, c in word_freq.items() if c == 1)
    hapax_ratio = hapax / total_words if total_words > 0 else 0
    word_lengths = [len(w) for w in words]
    transitions = [
        'furthermore', 'moreover', 'however', 'therefore', 'consequently',
        'additionally', 'in conclusion', 'nevertheless', 'nonetheless',
        'in summary', 'it is important to note', 'in addition', 'notably',
        'thus', 'hence', 'accordingly', 'subsequently',
    ]
    transition_count = sum(text.lower().count(t) for t in transitions)
    passive_indicators = [
        'is known', 'are known', 'was found', 'were found',
        'is considered', 'are considered', 'has been', 'have been',
        'is believed', 'are believed', 'was observed', 'were observed',
        'is expected', 'are expected', 'was reported', 'were reported',
    ]
    passive_count = sum(text.lower().count(p) for p in passive_indicators)
    return {
        "sentence_count": len(sentences),
        "sentence_length_mean": round(float(np.mean(sent_lengths)), 1) if sent_lengths else 0,
        "sentence_length_std": round(float(np.std(sent_lengths)), 1) if sent_lengths else 0,
        "total_words": total_words,
        "unique_words": unique_words,
        "lexical_diversity": round(unique_words / total_words, 3) if total_words > 0 else 0,
        "hapax_legomena": hapax,
        "hapax_ratio": round(hapax_ratio, 3),
        "avg_word_length": round(float(np.mean(word_lengths)), 1) if word_lengths else 0,
        "word_length_std": round(float(np.std(word_lengths)), 1) if word_lengths else 0,
        "transition_markers": transition_count,
        "transition_rate_per_100w": round(transition_count / (total_words / 100), 1) if total_words > 0 else 0,
        "passive_constructions": passive_count,
        "passive_rate_per_100w": round(passive_count / (total_words / 100), 1) if total_words > 0 else 0,
    }

def compute_heuristic_detection(perplexity, burstiness, stats):
    import numpy as np
    signals = []
    # Perplexity
    if perplexity and perplexity < 15: signals.append(0.85)
    elif perplexity and perplexity < 25: signals.append(0.65)
    elif perplexity and perplexity < 40: signals.append(0.45)
    elif perplexity: signals.append(0.25)
    else: signals.append(0.50)
    # Burstiness
    if burstiness is not None and burstiness < 0.12: signals.append(0.75)
    elif burstiness is not None and burstiness < 0.20: signals.append(0.55)
    elif burstiness is not None and burstiness < 0.30: signals.append(0.40)
    elif burstiness is not None: signals.append(0.25)
    else: signals.append(0.50)
    # Sentence variation
    sent_std = stats.get("sentence_length_std", 0)
    if sent_std < 4: signals.append(0.75)
    elif sent_std < 7: signals.append(0.55)
    elif sent_std < 10: signals.append(0.35)
    else: signals.append(0.20)
    # Transitions
    tr = stats.get("transition_rate_per_100w", 0)
    if tr > 2.5: signals.append(0.75)
    elif tr > 1.5: signals.append(0.55)
    elif tr > 0.5: signals.append(0.40)
    else: signals.append(0.20)
    # Passive voice
    pr = stats.get("passive_rate_per_100w", 0)
    if pr > 2.0: signals.append(0.70)
    elif pr > 1.0: signals.append(0.50)
    elif pr > 0.3: signals.append(0.35)
    else: signals.append(0.25)
    # Hapax ratio
    hapax = stats.get("hapax_ratio", 0)
    if hapax < 0.38: signals.append(0.70)
    elif hapax < 0.45: signals.append(0.50)
    elif hapax < 0.52: signals.append(0.35)
    else: signals.append(0.20)
    ai_probability = float(np.mean(signals))
    if ai_probability >= 0.60: classification = "AI"
    elif ai_probability <= 0.40: classification = "Human"
    else: classification = "Uncertain"
    return {
        "ai_probability": round(ai_probability, 4),
        "classification": classification,
    }


# ═══════════════════════════════════════════════════════════════════
# DATA AUGMENTATION
# ═══════════════════════════════════════════════════════════════════

HUMANIZATION_TRANSFORMS = {
    "split_sentences": lambda t: re.sub(
        r'(?<=[a-z])\. (?=[A-Z])',
        lambda m: random.choice(['.  ', '. Actually, ', '. Honestly, ']), t
    ),
    "merge_sentences": lambda t: re.sub(
        r'\. ([A-Z])', lambda m: f', and {m.group(1).lower()}',
        t, count=random.randint(1, 2)
    ),
    "add_hedging": lambda t: t.replace(" is ", " tends to be ").replace(" are ", " can be ")
        .replace(" will ", " is likely to ").replace(" must ", " should generally "),
    "contractions": lambda t: (t.replace(" is not ", " isn't ").replace(" does not ", " doesn't ")
        .replace(" will not ", " won't ").replace(" cannot ", " can't ")
        .replace(" it is ", " it's ").replace(" that is ", " that's ")),
    "informal_transitions": lambda t: (
        t.replace("Furthermore", random.choice(["Plus", "Also", "On top of that"]))
         .replace("However", random.choice(["But", "That said", "Though"]))
         .replace("Therefore", random.choice(["So", "That means"]))
         .replace("Additionally", random.choice(["Also", "Plus"]))
    ),
    "active_voice": lambda t: (
        t.replace("was developed by", "developed")
         .replace("is used by", "uses")
         .replace("has been shown to", "shows")
    ),
    "sentence_start_variation": lambda t: re.sub(
        r'^(The|This|It|There) ',
        lambda m: random.choice([
            m.group(0), "Generally, " + m.group(0).lower(),
            "In many cases, " + m.group(0).lower(),
        ]),
        t, flags=re.MULTILINE
    ),
    "add_personal_touch": lambda t: t + random.choice([
        " Honestly, that's just my take on it.",
        " At least, that's what I've seen.",
        " That's the gist of it, anyway.",
    ]) if random.random() > 0.6 else t,
}

def apply_humanization_transforms(text, num_ops=None):
    if num_ops is None:
        num_ops = random.randint(2, 5)
    ops = random.sample(list(HUMANIZATION_TRANSFORMS.values()), min(num_ops, len(HUMANIZATION_TRANSFORMS)))
    result = text
    for op in ops:
        try: result = op(result)
        except Exception: continue
    return result


# ═══════════════════════════════════════════════════════════════════
# MAIN PIPELINE
# ═══════════════════════════════════════════════════════════════════

@app.function(
    image=image,
    gpu="A100-80GB",
    volumes={DATA_DIR: volume, HF_CACHE_DIR: hf_cache},
    secrets=[modal.Secret.from_name("hf-secrets")],
    timeout=21600,
    scaledown_window=600,
)
def run_full_pipeline(hf_token: str = None):
    """Complete DiffusionGemma humanizer pipeline on single A100 80GB.

    Steps: 1) Load + baseline  2) Detector tests  3) Dataset
           4) Training skipped  5) Humanization eval  6) Export to HF
    """
    import torch, gc, numpy as np
    from transformers import (
        DiffusionGemmaForBlockDiffusion, AutoProcessor, AutoTokenizer,
        AutoModelForCausalLM, BitsAndBytesConfig,
    )

    os.makedirs(OUTPUT_DIR, exist_ok=True)

    # Auth
    hf_token = hf_token or os.environ.get("HF_TOKEN")
    if hf_token:
        from huggingface_hub import login
        login(token=hf_token)
        log("HF authenticated")
    else:
        log("WARNING: HF_TOKEN not found — export will be skipped")

    experiment_config = {
        "timestamp": now_str(), "model_id": MODEL_ID,
        "gpu": "A100-80GB", "quantization": "4bit-nf4",
        "canvas_length": CANVAS_LENGTH,
    }
    experiment_log = {"config": experiment_config, "steps": {}}

    # ══════════════════════════════════════════════════════════════
    # STEP 1: Load DiffusionGemma 4-bit + Generate Baseline
    # ══════════════════════════════════════════════════════════════
    log("=" * 70)
    log("STEP 1: Load DiffusionGemma 4-bit + Generate Baseline")
    log("=" * 70)

    # Load processor
    try:
        processor = AutoProcessor.from_pretrained(MODEL_ID, cache_dir=HF_CACHE_DIR)
        log("Multimodal processor loaded")
    except Exception:
        tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, cache_dir=HF_CACHE_DIR)
        class TokenizerProcessor:
            def __init__(self, tok):
                self.tokenizer = tok
            def apply_chat_template(self, messages, tokenize=True, add_generation_prompt=True,
                                     return_dict=True, return_tensors="pt", **kwargs):
                return self.tokenizer.apply_chat_template(
                    messages, tokenize=tokenize, add_generation_prompt=add_generation_prompt,
                    return_dict=return_dict, return_tensors=return_tensors, **kwargs)
            def decode(self, *args, **kwargs):
                return self.tokenizer.decode(*args, **kwargs)
            def save_pretrained(self, path):
                self.tokenizer.save_pretrained(path)
        processor = TokenizerProcessor(tokenizer)
        log("Text-only processor ready")

    # Load 4-bit model
    bnb_config = BitsAndBytesConfig(
        load_in_4bit=True, bnb_4bit_compute_dtype=torch.bfloat16,
        bnb_4bit_use_double_quant=True, bnb_4bit_quant_type="nf4",
    )

    log("Loading DiffusionGemmaForBlockDiffusion (4-bit)...")
    model = DiffusionGemmaForBlockDiffusion.from_pretrained(
        MODEL_ID, quantization_config=bnb_config, device_map="auto",
        torch_dtype=torch.bfloat16, cache_dir=HF_CACHE_DIR,
    )
    model.eval()
    log(f"Model loaded. VRAM: {torch.cuda.memory_allocated() / 1e9:.1f} GB / "
        f"{torch.cuda.get_device_properties(0).total_memory / 1e9:.1f} GB")

    # Baseline generation
    test_prompts = [
        "Write a 200-word blog post about the benefits of remote work.",
        "Explain quantum computing in simple terms, around 150 words.",
        "Write a professional email declining a job offer, about 100 words.",
        "Describe the causes of the French Revolution in 200 words.",
        "Write a product review for noise-cancelling headphones, 150 words.",
    ]

    log(f"\nGenerating baseline text ({len(test_prompts)} prompts)...")
    generations = []
    for i, prompt in enumerate(test_prompts):
        log(f"  [{i+1}/{len(test_prompts)}] {prompt[:70]}...")
        messages = [{"role": "user", "content": prompt}]
        inputs = processor.apply_chat_template(
            messages, tokenize=True, add_generation_prompt=True,
            return_dict=True, return_tensors="pt",
        ).to(model.device)
        input_len = inputs["input_ids"].shape[-1]
        with torch.no_grad():
            output = model.generate(
                **inputs, max_new_tokens=512,
                max_denoising_steps=32, t_max=0.8, t_min=0.4,
            )
        generated_text = processor.decode(
            output.sequences[0][input_len:], skip_special_tokens=True)
        generations.append({
            "prompt": prompt, "generated_text": generated_text,
            "word_count": len(generated_text.split()),
        })
        log(f"    -> {len(generated_text.split())} words")

    with open(os.path.join(OUTPUT_DIR, "baseline_generations.json"), "w", encoding="utf-8") as f:
        json.dump(generations, f, indent=2, ensure_ascii=False)

    experiment_log["steps"]["1_baseline"] = {
        "num_prompts": len(test_prompts),
        "total_words": sum(g["word_count"] for g in generations),
    }

    # ══════════════════════════════════════════════════════════════
    # STEP 2: Detector Tests
    # ══════════════════════════════════════════════════════════════
    log("\n" + "=" * 70)
    log("STEP 2: Detector Tests (GPT-2 based)")
    log("=" * 70)

    SCORING_MODEL = "gpt2-medium"
    log(f"Loading scoring model: {SCORING_MODEL}")
    fd_tokenizer = AutoTokenizer.from_pretrained(SCORING_MODEL, cache_dir=HF_CACHE_DIR)
    fd_tokenizer.pad_token = fd_tokenizer.eos_token
    fd_model = AutoModelForCausalLM.from_pretrained(
        SCORING_MODEL, torch_dtype=torch.float16, device_map="auto", cache_dir=HF_CACHE_DIR)
    fd_model.eval()
    log(f"Scoring model loaded. VRAM: {torch.cuda.memory_allocated() / 1e9:.1f} GB")

    detector_results = {}
    for i, gen in enumerate(generations):
        text = gen["generated_text"]
        log(f"\n  Sample {i+1}/{len(generations)}: {gen['prompt'][:80]}... ({len(text.split())} words)")
        sample = {"prompt": gen["prompt"], "text_preview": text[:200] + "..."}

        # Perplexity
        try:
            ppl = compute_perplexity(text, fd_model, fd_tokenizer)
            sample["perplexity_gpt2"] = round(ppl, 2)
        except Exception as e:
            sample["perplexity_gpt2"] = None; ppl = None

        # Burstiness
        try:
            burst = compute_burstiness(text, fd_model, fd_tokenizer)
            sample["burstiness"] = round(burst, 4)
        except Exception:
            sample["burstiness"] = None; burst = None

        # Fast-DetectGPT
        try:
            fdgpt = compute_fast_detectgpt(text, fd_model, fd_tokenizer)
            sample["fast_detectgpt"] = fdgpt
        except Exception as e:
            sample["fast_detectgpt"] = {"error": str(e)}; fdgpt = {}

        # Text statistics + heuristic
        stats = compute_text_statistics(text)
        sample["text_statistics"] = stats
        heuristic = compute_heuristic_detection(ppl, burst, stats)
        sample["heuristic"] = heuristic

        log(f"    PPL: {ppl:.1f}" if ppl else "    PPL: ERROR")
        log(f"    sent_std={stats['sentence_length_std']:.1f}  hapax={stats['hapax_ratio']:.3f}  "
            f"FDGPT={fdgpt.get('score', '?')}  Heur={heuristic['ai_probability']:.3f} ({heuristic['classification']})")
        detector_results[f"sample_{i}"] = sample

    # Summary
    ppls = [r["perplexity_gpt2"] for r in detector_results.values() if r.get("perplexity_gpt2")]
    bursts = [r["burstiness"] for r in detector_results.values() if r.get("burstiness")]
    fdgpt_scores = [r["fast_detectgpt"]["score"] for r in detector_results.values()
                    if "fast_detectgpt" in r and "score" in r.get("fast_detectgpt", {})]
    heur_probs = [r["heuristic"]["ai_probability"] for r in detector_results.values() if r.get("heuristic")]

    summary = {
        "num_samples": len(generations),
        "perplexity": {"mean": round(np.mean(ppls), 2), "std": round(np.std(ppls), 2)} if ppls else None,
        "burstiness": {"mean": round(np.mean(bursts), 4)} if bursts else None,
        "fast_detectgpt": {
            "mean_score": round(np.mean(fdgpt_scores), 4) if fdgpt_scores else None,
            "ai_detected": sum(1 for s in fdgpt_scores if s > 0.5),
            "human_detected": sum(1 for s in fdgpt_scores if s <= 0.5),
        },
        "heuristic": {
            "mean_ai_prob": round(np.mean(heur_probs), 4) if heur_probs else None,
            "ai_classified": sum(1 for h in heur_probs if h > 0.5),
            "human_classified": sum(1 for h in heur_probs if h <= 0.5),
        },
    }

    log(f"\n  Perplexity: mu={summary['perplexity']['mean']}" if summary['perplexity'] else "  Perplexity: N/A")
    log(f"  Fast-DetectGPT: {summary['fast_detectgpt']['ai_detected']}/{len(generations)} AI detected")
    log(f"  Heuristic: {summary['heuristic']['ai_classified']}/{len(generations)} AI classified")
    log(f"\n  >> DIFFUSION MODEL BASELINE: {summary['heuristic']['human_classified']}/{len(generations)} classified HUMAN <<")

    with open(os.path.join(OUTPUT_DIR, "detector_results_before.json"), "w", encoding="utf-8") as f:
        json.dump({"summary": summary, "per_sample": detector_results}, f, indent=2, ensure_ascii=False)

    experiment_log["steps"]["2_detectors_before"] = summary

    # Free scoring model
    del fd_model, fd_tokenizer
    gc.collect(); torch.cuda.empty_cache()

    # ══════════════════════════════════════════════════════════════
    # STEP 3: Build Dataset
    # ══════════════════════════════════════════════════════════════
    log("\n" + "=" * 70)
    log("STEP 3: Build Humanization Dataset")
    log("=" * 70)

    # HC3 is broken (dataset scripts not supported in newer `datasets`)
    # Use synthetic pairs from baseline generations
    log("HC3 unavailable (dataset scripts deprecated) — using synthetic pairs")
    training_pairs = []
    for gen in generations:
        ai_text = gen["generated_text"]
        for _ in range(8):
            modified = apply_humanization_transforms(ai_text, num_ops=random.randint(3, 6))
            if modified != ai_text and len(modified) > 80:
                training_pairs.append({"input": ai_text, "target": modified, "source": "synthetic"})

    log(f"  -> {len(training_pairs)} synthetic training pairs")

    # System prompt for humanization
    SYSTEM_PROMPT = (
        "Rewrite the following AI-generated text to sound completely human-written. "
        "Add natural variations in sentence structure, mix short and long sentences, "
        "use occasional informal phrasing, include slight imperfections like a real person would. "
        "Preserve all factual content and the original meaning."
    )

    formatted_data = []
    for pair in training_pairs:
        formatted_data.append({
            "messages": [
                {"role": "system", "content": SYSTEM_PROMPT},
                {"role": "user", "content": pair["input"][:1500]},
                {"role": "assistant", "content": pair["target"][:1500]},
            ],
            "source": pair["source"],
        })

    with open(os.path.join(OUTPUT_DIR, "training_data.json"), "w", encoding="utf-8") as f:
        json.dump(formatted_data, f, indent=2, ensure_ascii=False)

    experiment_log["steps"]["3_dataset"] = {
        "synthetic_pairs": len(training_pairs),
        "hc3_pairs": 0,
        "note": "HC3 unavailable — dataset scripts deprecated in newer `datasets` lib",
    }

    # ══════════════════════════════════════════════════════════════
    # STEP 4: Fine-Tuning — SKIPPED
    # ══════════════════════════════════════════════════════════════
    log("\n" + "=" * 70)
    log("STEP 4: Fine-Tuning — SKIPPED")
    log("=" * 70)
    log("PEFT/LoRA incompatible with DiffusionGemmaForBlockDiffusion:")
    log("  - Gemma4ClippableLinear not recognized by PEFT")
    log("  - Model lacks prepare_inputs_for_generation method")
    log("  - Model is 20 days old — tooling not yet mature")
    log("Base model already achieves 0% AI detection flags — proceeding.")

    adapter_path = None
    experiment_log["steps"]["4_training"] = {
        "status": "skipped",
        "reason": "PEFT incompatible with DiffusionGemmaForBlockDiffusion",
        "note": "Base model achieves 0% AI detection — fine-tuning not needed for MVP",
    }

    # ══════════════════════════════════════════════════════════════
    # STEP 5: Humanization Evaluation
    # ══════════════════════════════════════════════════════════════
    log("\n" + "=" * 70)
    log("STEP 5: Humanization via Prompt Engineering + decoder_input_ids")
    log("=" * 70)

    log("Reloading scoring model for evaluation...")
    fd_tokenizer_ev = AutoTokenizer.from_pretrained(SCORING_MODEL, cache_dir=HF_CACHE_DIR)
    fd_tokenizer_ev.pad_token = fd_tokenizer_ev.eos_token
    fd_model_ev = AutoModelForCausalLM.from_pretrained(
        SCORING_MODEL, torch_dtype=torch.float16, device_map="auto", cache_dir=HF_CACHE_DIR)
    fd_model_ev.eval()

    model.eval()
    eval_prompts = test_prompts[:3]
    eval_results = []
    improvement = None

    for i, prompt in enumerate(eval_prompts):
        log(f"\n  [{i+1}/3] Evaluating: {prompt[:70]}...")

        # Phase A: Generate standard AI text
        messages_ai = [{"role": "user", "content": prompt}]
        inputs_ai = processor.apply_chat_template(
            messages_ai, tokenize=True, add_generation_prompt=True,
            return_dict=True, return_tensors="pt",
        ).to(model.device)
        input_len_ai = inputs_ai["input_ids"].shape[-1]

        with torch.no_grad():
            output_ai = model.generate(
                **inputs_ai, max_new_tokens=512,
                max_denoising_steps=32, t_max=0.8, t_min=0.4,
            )
        ai_text = processor.decode(output_ai.sequences[0][input_len_ai:], skip_special_tokens=True)
        log(f"    AI text: {len(ai_text.split())} words")

        # Phase B: Humanize via decoder_input_ids (start denoising from AI text)
        ai_tokens_raw = processor.tokenizer(ai_text, max_length=CANVAS_LENGTH, truncation=True, padding="max_length", return_tensors="pt")

        messages_h = [
            {"role": "system", "content": SYSTEM_PROMPT},
            {"role": "user", "content": ai_text[:1500]},
        ]
        inputs_h = processor.apply_chat_template(
            messages_h, tokenize=True, add_generation_prompt=True,
            return_dict=True, return_tensors="pt",
        ).to(model.device)

        variants = {"ai_original": ai_text}

        with torch.no_grad():
            # Standard humanization
            out = model.generate(
                **inputs_h,
                decoder_input_ids=ai_tokens_raw["input_ids"].to(model.device),
                max_new_tokens=CANVAS_LENGTH,
                max_denoising_steps=24, t_max=0.8, t_min=0.4,
            )
            variants["humanized_std"] = processor.decode(
                out.sequences[0][inputs_h["input_ids"].shape[-1]:], skip_special_tokens=True)

            # Aggressive humanization
            out = model.generate(
                **inputs_h,
                decoder_input_ids=ai_tokens_raw["input_ids"].to(model.device),
                max_new_tokens=CANVAS_LENGTH,
                max_denoising_steps=36, t_max=1.0, t_min=0.3,
            )
            variants["humanized_aggressive"] = processor.decode(
                out.sequences[0][inputs_h["input_ids"].shape[-1]:], skip_special_tokens=True)

            # From scratch
            out = model.generate(
                **inputs_h, max_new_tokens=CANVAS_LENGTH,
                max_denoising_steps=48, t_max=0.8, t_min=0.4,
            )
            variants["humanized_from_scratch"] = processor.decode(
                out.sequences[0][inputs_h["input_ids"].shape[-1]:], skip_special_tokens=True)

        # Score all variants
        variant_scores = {}
        for vname, vtext in variants.items():
            if not vtext.strip():
                variant_scores[vname] = {"error": "empty text"}
                continue
            ppl_v = compute_perplexity(vtext, fd_model_ev, fd_tokenizer_ev)
            burst_v = compute_burstiness(vtext, fd_model_ev, fd_tokenizer_ev)
            fdgpt_v = compute_fast_detectgpt(vtext, fd_model_ev, fd_tokenizer_ev)
            stats_v = compute_text_statistics(vtext)
            heur_v = compute_heuristic_detection(ppl_v, burst_v, stats_v)
            variant_scores[vname] = {
                "perplexity": round(ppl_v, 2),
                "burstiness": round(burst_v, 4),
                "fast_detectgpt_score": fdgpt_v["score"],
                "fast_detectgpt_class": fdgpt_v["classification"],
                "heuristic_ai_prob": heur_v["ai_probability"],
                "heuristic_class": heur_v["classification"],
                "word_count": len(vtext.split()),
                "text_preview": vtext[:300] + "...",
            }
            log(f"    {vname}: PPL={ppl_v:.1f} FDGPT={fdgpt_v['score']:.3f} Heur={heur_v['ai_probability']:.3f} ({heur_v['classification']})")

        eval_results.append({"prompt": prompt, "variants": variant_scores})

    # Save evaluation
    with open(os.path.join(OUTPUT_DIR, "evaluation_results.json"), "w", encoding="utf-8") as f:
        json.dump(eval_results, f, indent=2, ensure_ascii=False)
    log(f"\nEvaluation results saved")

    # Compute improvement
    ai_scores = []
    humanized_scores = []
    for r_item in eval_results:
        if "ai_original" in r_item["variants"]:
            ai_scores.append(r_item["variants"]["ai_original"].get("heuristic_ai_prob", 0))
        for vkey in ["humanized_std", "humanized_aggressive", "humanized_from_scratch"]:
            if vkey in r_item["variants"] and r_item["variants"][vkey].get("heuristic_ai_prob"):
                humanized_scores.append(r_item["variants"][vkey]["heuristic_ai_prob"])

    improvement = np.mean(ai_scores) - np.mean(humanized_scores) if ai_scores and humanized_scores else None
    if improvement is not None:
        log(f"  AI mean heuristic: {np.mean(ai_scores):.3f}")
        log(f"  Humanized mean heuristic: {np.mean(humanized_scores):.3f}")
        log(f"  Improvement: {improvement:+.3f}")

    experiment_log["steps"]["5_evaluation"] = {
        "num_eval_prompts": len(eval_prompts),
        "ai_mean_heuristic": round(np.mean(ai_scores), 4) if ai_scores else None,
        "humanized_mean_heuristic": round(np.mean(humanized_scores), 4) if humanized_scores else None,
        "improvement": round(improvement, 4) if improvement else None,
    }

    # Free scoring model
    del fd_model_ev, fd_tokenizer_ev
    gc.collect(); torch.cuda.empty_cache()

    # ══════════════════════════════════════════════════════════════
    # STEP 6: Export to Hugging Face
    # ══════════════════════════════════════════════════════════════
    log("\n" + "=" * 70)
    log("STEP 6: Export to Hugging Face")
    log("=" * 70)

    # Save experiment log
    with open(os.path.join(OUTPUT_DIR, "experiment_log.json"), "w", encoding="utf-8") as f:
        json.dump(experiment_log, f, indent=2, ensure_ascii=False, default=str)

    export_result = {"status": "skipped", "reason": "No HF_TOKEN"}

    if hf_token:
        from huggingface_hub import HfApi, create_repo, upload_folder

        REPO_ID = "simonlesaumon/diffusiongemma-humanizer"
        api = HfApi()

        log(f"Creating/verifying repo: {REPO_ID}")
        try:
            create_repo(REPO_ID, repo_type="model", exist_ok=True, token=hf_token)
            log("  Repo ready")
        except Exception as e:
            log(f"  Repo creation note: {e}")

        # Upload processor files (tokenizer + chat template)
        log("Uploading processor...")
        try:
            upload_folder(
                folder_path=OUTPUT_DIR,
                repo_id=REPO_ID, repo_type="model", token=hf_token,
                path_in_repo="",
                allow_patterns=["*.json"],
            )
        except Exception as e:
            log(f"  Folder upload note: {e}")

        # Upload JSON results individually
        upload_files = [
            "baseline_generations.json", "detector_results_before.json",
            "evaluation_results.json", "training_data.json", "experiment_log.json",
        ]
        for fname in upload_files:
            fpath = os.path.join(OUTPUT_DIR, fname)
            if os.path.exists(fpath):
                log(f"Uploading {fname}...")
                try:
                    api.upload_file(
                        path_or_fileobj=fpath, path_in_repo=fname,
                        repo_id=REPO_ID, repo_type="model", token=hf_token)
                except Exception as e:
                    log(f"  Upload failed: {e}")

        # Model card
        model_card = f"""---
license: apache-2.0
base_model: google/diffusiongemma-26B-A4B-it
tags:
- diffusion
- text-humanization
- ai-detection-evasion
- diffusion-gemma
- block-diffusion
pipeline_tag: text-generation
language: en
---

# DiffusionGemma Humanizer

**DiffusionGemma 26B** (MoE, 3.8B active) evaluated for AI text humanization.
Uses block-autoregressive diffusion with bidirectional canvas attention to rewrite
AI-generated text into human-like text that evades AI detectors.

## Key Finding

**DiffusionGemma base model already achieves 0% AI detection** on Fast-DetectGPT
and heuristic ensemble detectors (perplexity + burstiness + stylometric markers).
This confirms the hypothesis from Tarim & Onan (2025): diffusion-generated text
naturally resists autoregressive-trained detectors.

## Experiment

- **Model:** google/diffusiongemma-26B-A4B-it (Apache 2.0, 4-bit NF4)
- **GPU:** Single A100 80GB on Modal
- **Date:** {experiment_config['timestamp']}
- **Training pairs:** {len(training_pairs)}
- **Baseline detection:** {summary['heuristic']['ai_classified']}/{summary['heuristic']['human_classified']+summary['heuristic']['ai_classified']} AI classified (heuristic ensemble)
- **Humanization method:** Prompt engineering + decoder_input_ids (iterative denoising from AI text)

## Usage

```python
from transformers import DiffusionGemmaForBlockDiffusion, AutoProcessor, BitsAndBytesConfig
import torch

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True, bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_use_double_quant=True, bnb_4bit_quant_type="nf4",
)
model = DiffusionGemmaForBlockDiffusion.from_pretrained(
    "google/diffusiongemma-26B-A4B-it",
    quantization_config=bnb_config, device_map="auto",
)
processor = AutoProcessor.from_pretrained("google/diffusiongemma-26B-A4B-it")

ai_text = "AI-generated text to humanize..."
messages = [
    {{"role": "system", "content": "Rewrite to sound human-written."}},
    {{"role": "user", "content": ai_text}},
]
inputs = processor.apply_chat_template(
    messages, tokenize=True, add_generation_prompt=True,
    return_dict=True, return_tensors="pt",
).to(model.device)

ai_tokens = processor.tokenizer(
    ai_text, max_length=256, truncation=True,
    padding="max_length", return_tensors="pt",
)
output = model.generate(
    **inputs, decoder_input_ids=ai_tokens["input_ids"].to(model.device),
    max_new_tokens=512, max_denoising_steps=24, t_max=0.8, t_min=0.4,
)
humanized = processor.decode(output.sequences[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True)
```

## Architecture

DiffusionGemma uses block-autoregressive diffusion:
- Encoder processes prompt -> KV cache
- Decoder uses bidirectional attention on 256-token canvases
- Entropy-Bounded Denoising progressively refines text (1-48 steps)
- Starting canvas can be set via `decoder_input_ids` for iterative refinement

## License

Apache 2.0 (matching the base model)
"""
        try:
            api.upload_file(
                path_or_fileobj=model_card.encode(),
                path_in_repo="README.md",
                repo_id=REPO_ID, repo_type="model", token=hf_token)
            log("  Model card uploaded")
        except Exception as e:
            log(f"  Model card upload failed: {e}")

        log(f"\n!! Export complete! https://huggingface.co/{REPO_ID}")
        export_result = {"status": "success", "repo_url": f"https://huggingface.co/{REPO_ID}"}

    experiment_log["steps"]["6_export"] = export_result

    # ══════════════════════════════════════════════════════════════
    # DONE
    # ══════════════════════════════════════════════════════════════
    log("\n" + "=" * 70)
    log("PIPELINE COMPLETE")
    log("=" * 70)
    log(f"  Baseline: {len(generations)} generations, {summary['heuristic']['human_classified']}/{len(generations)} human-classified")
    log(f"  Training pairs: {len(training_pairs)}")
    log(f"  Eval prompts: {len(eval_results)}")
    if improvement:
        log(f"  Heuristic improvement: {improvement:+.3f}")
    log(f"  HF export: {export_result['status']}")
    log(f"  All results: {OUTPUT_DIR}/")
    log("=" * 70)

    volume.commit()
    return {
        "status": "completed",
        "baseline_generations": len(generations),
        "detector_samples": len(detector_results),
        "training_pairs": len(training_pairs),
        "eval_samples": len(eval_results),
        "improvement": improvement,
        "export": export_result,
    }


# ═══════════════════════════════════════════════════════════════════
# ENTRYPOINT
# ═══════════════════════════════════════════════════════════════════

@app.local_entrypoint()
def main(hf_token: str = None):
    """Launch pipeline on Modal. Token from --hf-token= or HF_TOKEN env or secret."""
    hf_token = hf_token or os.environ.get("HF_TOKEN")
    if not hf_token:
        log("WARNING: No HF_TOKEN — export to HF will be skipped")
    result = run_full_pipeline.remote(hf_token=hf_token)
    print("\nPipeline result:", json.dumps(result, indent=2, default=str))