Upload README.md with huggingface_hub

README.md

@@ -7,88 +7,51 @@ tags:
 - ai-detection-evasion
 - diffusion-gemma
 - block-diffusion
-- research
 pipeline_tag: text-generation
 language: en
 ---
 
 # DiffusionGemma Humanizer
 
-**Research, code, and pipeline** for humanizing AI-generated text using
-Google's **DiffusionGemma 26B** (MoE, 3.8B active parameters, Apache 2.0).
+**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 achieves ~0% AI detection** on Fast-DetectGPT +
-heuristic ensemble detectors (perplexity + burstiness + stylometric markers).
-Diffusion-generated text naturally resists autoregressive-trained detectors,
-confirming the hypothesis from Tarim & Onan (2025).
+**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.
 
-## How It Works
+## Experiment
 
-DiffusionGemma uses **block-autoregressive diffusion** with bidirectional
-canvas attention. Instead of generating text left-to-right like GPT models,
-it iteratively denoises 256-token blocks, producing text with fundamentally
-different statistical fingerprints that evade standard AI detectors.
+- **Model:** google/diffusiongemma-26B-A4B-it (Apache 2.0, 4-bit NF4)
+- **GPU:** Single A100 80GB on Modal
+- **Date:** 20260629-201308
+- **Training pairs:** 39
+- **Baseline detection:** 0/5 AI classified (heuristic ensemble)
+- **Humanization method:** Prompt engineering + decoder_input_ids (iterative denoising from AI text)
 
-The humanization pipeline:
-1. Load DiffusionGemma 26B in 4-bit (NF4) on A100 80GB via Modal
-2. Generate baseline text with entropy-bounded denoising
-3. Test against open-source detectors (perplexity, burstiness, Fast-DetectGPT)
-4. Humanize via prompt engineering + decoder_input_ids (start denoising from AI text)
-5. Export results to this repository
-
-## Repository Contents
-
-- `research_report.md` — Gemma models, diffusion LMs, Modal pricing
-- `research_datasets_training.md` — Training data survey (HC3, RAID, M4, etc.)
-- `commercial_ai_detectors_report.md` — Pangram, GPTZero, Originality.ai analysis
-- `research/architecture-strategy.md` — Architecture decisions & cost breakdown
-- `research/technical-diffusion-text-humanization-2026-06-29.md` — Full lit review
-- `modal_project/app.py` — Complete 6-step Modal pipeline
-- `scripts/` — Pipeline launchers
-
-## 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
-
-## Why Diffusion for Humanization?
-
-1. **Different statistical signature** — Diffusion-generated text follows different token probability distributions than autoregressive text, making it OOD for AR-trained detectors
-2. **Bidirectional attention** — Holistic text understanding enables coherent style manipulation
-3. **Iterative denoising** — Progressive refinement from AI text toward human-like output
-4. **No left-to-right bias** — Avoids the "reversal curse" and other AR artifacts
-
-## Humanization Code Example
+## Usage
 
 ```python
 from transformers import DiffusionGemmaForBlockDiffusion, AutoProcessor, BitsAndBytesConfig
 import torch
 
-bn_config = BitsAndBytesConfig(
-    load_in_4bit=True,
-    bnb_4bit_compute_dtype=torch.bfloat16,
-    bnb_4bit_use_double_quant=True,
-    bnb_4bit_quant_type="nf4",
+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=bn_config,
-    device_map="auto",
+    quantization_config=bnb_config, device_map="auto",
 )
 processor = AutoProcessor.from_pretrained("google/diffusiongemma-26B-A4B-it")
 
-# Starting canvas = AI text (not random)
 ai_text = "AI-generated text to humanize..."
-ai_tokens = processor.tokenizer(ai_text, max_length=256, truncation=True,
-                                 padding="max_length", return_tensors="pt")
-
 messages = [
-    {"role": "system", "content": "Rewrite to sound completely human-written."},
+    {"role": "system", "content": "Rewrite to sound human-written."},
     {"role": "user", "content": ai_text},
 ]
 inputs = processor.apply_chat_template(
@@ -96,29 +59,25 @@ inputs = processor.apply_chat_template(
     return_dict=True, return_tensors="pt",
 ).to(model.device)
 
-# Iterative denoising from AI text toward human-like output
-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,
+ai_tokens = processor.tokenizer(
+    ai_text, max_length=256, truncation=True,
+    padding="max_length", return_tensors="pt",
 )
-humanized = processor.decode(
-    output.sequences[0][inputs["input_ids"].shape[-1]:],
-    skip_special_tokens=True,
+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)
 ```
 
-## References
+## Architecture
 
-- [DiffusionGemma Model Card](https://huggingface.co/google/diffusiongemma-26B-A4B-it)
-- Tarım & Onan (2025) — "Can You Detect the Difference? Diffusion vs AR Text"
-- Sadasivan et al. (2023) — "Can AI-Generated Text be Reliably Detected?"
-- Cheng et al. (2025) — "Adversarial Paraphrasing for Humanizing AI Text"
-- [Modal GPU Pricing](https://modal.com/pricing)
+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).
+Apache 2.0 (matching the base model)