Upload README.md with huggingface_hub

README.md

@@ -0,0 +1,282 @@
+---
+license: apache-2.0
+base_model: zai-org/GLM-4.6V-Flash
+model_name: Elbaz-GLM-4.6V-Flash-PRISM
+tags:
+  - abliteration
+  - prism
+  - vision-language-model
+  - vlm
+  - glm
+  - gguf
+  - quantized
+language:
+  - en
+library_name: transformers
+pipeline_tag: image-text-to-text
+---
+
+<p align="center">
+  <img src="https://raw.githubusercontent.com/zai-org/GLM-V/refs/heads/main/resources/logo.svg" width="400"/>
+</p>
+
+# ELBAZ GLM-4.6V-FLASH PRISM
+
+**GLM-4.6V-Flash: A 10B Dense Vision-Language Model**
+
+[GLM-4.6V-Flash](https://huggingface.co/zai-org/GLM-4.6V-Flash) | [ZhipuAI](https://www.zhipuai.cn/)
+
+## Introduction
+
+**GLM-4.6V-Flash** is a 10.29B parameter dense Vision-Language Model (VLM) with a 40-layer transformer architecture and integrated vision encoder, capable of understanding both text and images.
+
+## Model Description
+
+This model is an **abliterated** version of [zai-org/GLM-4.6V-Flash](https://huggingface.co/zai-org/GLM-4.6V-Flash) that has had its refusal mechanisms removed using **PRISM (Projected Refusal Isolation via Subspace Modification)**. The model will respond to prompts that the original model would refuse.
+
+**Key Specs:**
+- 10.29B parameter dense Vision-Language Model
+- 40-layer transformer architecture
+- Integrated vision encoder for image understanding
+- 128K context length
+- Supports text, image, and video inputs
+
+### Motivation
+
+This project exists as **research and development experimentation** into understanding how large language models encode and enforce refusal behaviors, contributing to broader AI safety research by providing empirical data on refusal mechanism localization and tradeoffs between safety and capability.
+
+### Author
+
+**Eric Elbaz (Ex0bit)**
+
+## Model Tree
+
+```
+zai-org/GLM-4.6V-Flash (Base Model - BF16)
+└── Ex0bit/Elbaz-GLM-4.6V-Flash-PRISM (This Model)
+    └── Elbaz-GLM-4.6V-Flash-PRISM-IQ4_XS.gguf
+```
+
+## Available Quantizations
+
+| Quantization | Size | Description |
+|-------------|------|-------------|
+| IQ4_XS | 5.0 GB | Importance-weighted 4-bit, excellent quality |
+
+The IQ4_XS quantization uses importance-weighted quantization which provides better quality than standard Q4 quantizations at similar sizes. Embedding and output layers use Q6_K precision for optimal quality.
+
+## Prompt Format
+
+This model uses the GLM chat format with optional thinking/reasoning support:
+
+```
+[gMASK]<sop><|system|>
+{system_prompt}<|user|>
+{user_prompt}<|assistant|>
+```
+
+### Template Structure
+
+| Component | Token/Format |
+|-----------|-------------|
+| System Start | `<\|system\|>` |
+| User Start | `<\|user\|>` |
+| Assistant Start | `<\|assistant\|>` |
+| Thinking Start | `<think>` |
+| Thinking End | `</think>` |
+| End of Text | `<\|endoftext\|>` |
+
+### Special Tokens
+
+| Token | ID | Purpose |
+|-------|-----|---------|
+| `<\|system\|>` | 151335 | System prompt marker |
+| `<\|user\|>` | 151336 | User message marker |
+| `<\|assistant\|>` | 151337 | Assistant response marker |
+| `<think>` | 151350 | Reasoning block start |
+| `</think>` | 151351 | Reasoning block end |
+| `<\|endoftext\|>` | 151329 | EOS token |
+| `<\|begin_of_image\|>` | 151339 | Image input start |
+| `<\|end_of_image\|>` | 151340 | Image input end |
+
+## Technical Details
+
+### Performance Impact
+
+| Metric | Result |
+|--------|--------|
+| Refusal Bypass Rate | 100% |
+| English Output Rate | 100% |
+| KL Divergence | 0.0000 (no capability degradation) |
+| Response Coherence | Detailed, technically accurate |
+
+Testing shows that PRISM abliteration maintains full model coherence with no measurable capability degradation.
+
+## Quick Start
+
+### Using with llama.cpp
+
+```bash
+# Download the model
+huggingface-cli download Ex0bit/Elbaz-GLM-4.6V-Flash-PRISM \
+    Elbaz-GLM-4.6V-Flash-PRISM-IQ4_XS.gguf \
+    --local-dir .
+
+# Run inference
+./llama-cli -m Elbaz-GLM-4.6V-Flash-PRISM-IQ4_XS.gguf \
+    -p "[gMASK]<sop><|system|>
+You are a helpful assistant. You MUST respond in English only.<|user|>
+Your prompt here<|assistant|>
+" \
+    -n 2048 \
+    --temp 0.7 \
+    -ngl 999
+```
+
+### llama.cpp with llama-server
+
+```bash
+# Start the server
+./llama-server -m Elbaz-GLM-4.6V-Flash-PRISM-IQ4_XS.gguf \
+    --host 0.0.0.0 \
+    --port 8080 \
+    -ngl 999 \
+    -c 32768
+
+# Example API call
+curl http://localhost:8080/v1/chat/completions \
+    -H "Content-Type: application/json" \
+    -d '{
+        "messages": [
+            {"role": "system", "content": "You are a helpful assistant. You MUST respond in English only."},
+            {"role": "user", "content": "Your prompt here"}
+        ],
+        "temperature": 0.7
+    }'
+```
+
+### Using with Ollama
+
+```bash
+# Pull and run directly from Hugging Face
+ollama pull hf.co/Ex0bit/Elbaz-GLM-4.6V-Flash-PRISM
+ollama run hf.co/Ex0bit/Elbaz-GLM-4.6V-Flash-PRISM
+```
+
+> **Note:** The `hf.co/` prefix is required to pull from Hugging Face. Requires Ollama 0.3.0+.
+
+### Using with Transformers (Full Weights)
+
+```python
+from transformers import AutoModelForCausalLM, AutoProcessor
+
+model_id = "Ex0bit/Elbaz-GLM-4.6V-Flash-PRISM"
+processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)
+model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto", trust_remote_code=True)
+
+messages = [
+    {"role": "system", "content": [{"type": "text", "text": "You are a helpful assistant. You MUST respond in English only."}]},
+    {"role": "user", "content": [{"type": "text", "text": "Your prompt here"}]}
+]
+
+inputs = processor.apply_chat_template(
+    messages,
+    tokenize=True,
+    add_generation_prompt=True,
+    return_dict=True,
+    return_tensors="pt"
+).to(model.device)
+
+outputs = model.generate(**inputs, max_new_tokens=2048, temperature=0.7, do_sample=True)
+print(processor.decode(outputs[0], skip_special_tokens=False))
+```
+
+## PRISM Methodology
+
+### Method: Projected Refusal Isolation via Subspace Modification
+
+The model was abliterated using **PRISM** - a state-of-the-art abliteration methodology combining multiple principled techniques for effective refusal removal while preserving model capabilities.
+
+**Core Approach:**
+
+1. **Per-Layer Refusal Direction** - Each layer gets its own unique refusal direction (`r = harmful_mean - harmless_mean`) instead of a single global direction
+2. **Projected Direction Isolation** - Projects refusal direction orthogonal to harmless subspace to avoid "helpfulness confound"
+3. **Dynamic Layer-Wise Weight Kernel** - Bell curve distribution focusing on middle layers where refusal is encoded (weights range 0.45 to 2.24)
+4. **Winsorization** - Clips extreme values for numerical stability
+5. **KL Divergence Preservation** - Maintains 0.0000 KL divergence across all layers
+
+**Key Innovation:** Per-layer refusal directions preserve layer-specific behavior better than global averaging approaches.
+
+## Hardware Requirements
+
+| Quantization | Min RAM/VRAM | Recommended | Hardware Examples |
+|-------------|--------------|-------------|-------------------|
+| IQ4_XS | 8 GB | 12+ GB | RTX 3060 12GB, RTX 4070, Apple M1/M2/M3/M4 |
+
+### Tested Configurations
+
+| Hardware | RAM/VRAM | Status |
+|----------|----------|--------|
+| NVIDIA RTX GPU | 12+ GB | Works |
+| Apple Silicon | 16+ GB Unified | Works |
+
+**Note:** This is a relatively lightweight model that can run on consumer hardware with 12GB+ VRAM.
+
+## Vision Capabilities
+
+GLM-4.6V-Flash supports multimodal inputs:
+
+- **Images**: Use `<|begin_of_image|><|image|><|end_of_image|>` tags
+- **Videos**: Use `<|begin_of_video|><|video|><|end_of_video|>` tags
+
+Example with image:
+```python
+messages = [
+    {
+        "role": "user",
+        "content": [
+            {"type": "image", "image": "path/to/image.jpg"},
+            {"type": "text", "text": "What is in this image?"}
+        ]
+    }
+]
+```
+
+## Ethical Considerations
+
+This model has been modified to reduce safety guardrails. Users are responsible for:
+
+- Complying with all applicable laws and regulations
+- Not using the model for illegal activities
+- Understanding the potential risks of unrestricted AI responses
+- Implementing appropriate safeguards in production environments
+
+## License
+
+Apache 2.0 (same as base model [zai-org/GLM-4.6V-Flash](https://huggingface.co/zai-org/GLM-4.6V-Flash))
+
+## Citation
+
+```bibtex
+@misc{elbaz2025glm46vprism,
+  author = {Elbaz, Eric},
+  title = {Elbaz-GLM-4.6V-Flash-PRISM: An Abliterated GLM-4.6V Vision-Language Model},
+  year = {2025},
+  publisher = {Hugging Face},
+  howpublished = {\url{https://huggingface.co/Ex0bit/Elbaz-GLM-4.6V-Flash-PRISM}}
+}
+```
+
+## Acknowledgments
+
+- [ZhipuAI](https://www.zhipuai.cn/) for GLM-4.6V-Flash
+- [llama.cpp](https://github.com/ggerganov/llama.cpp) for quantization tools
+
+## Related Models
+
+- [zai-org/GLM-4.6V-Flash](https://huggingface.co/zai-org/GLM-4.6V-Flash) - Base model
+- [Ex0bit/Elbaz-Prime-Intellect-3_Prism_Abliterated](https://huggingface.co/Ex0bit/Elbaz-Prime-Intellect-3_Prism_Abliterated) - INTELLECT-3 abliterated
+
+---
+
+**Created by: Ex0bit (Eric Elbaz)**