Update README.md

README.md

@@ -23,24 +23,44 @@ tags:
 
 # **Qwen3-VL-4B-Instruct-Unredacted-MAX**
 
-> **Qwen3-VL-4B-Instruct-Unredacted-MAX** is an unredacted evolution built on top of **Qwen3-VL-4B-Instruct**. This model applies advanced abliterated training strategies designed to minimize internal refusal behaviors while preserving the core multimodal reasoning strengths of the original architecture. The result is a highly capable 4B vision-language model optimized for unrestricted, detailed reasoning and captioning across complex visual inputs.
+> **Qwen3-VL-4B-Instruct-Unredacted-MAX** is an optimized release built on top of **huihui-ai/Qwen3-VL-4B-Instruct-abliterated**. This version focuses on **improved packaging, updated Transformers compatibility, and stable multimodal inference behavior**, while preserving the core vision-language reasoning capabilities of the original architecture. The result is a capable **4B vision-language model** designed for efficient deployment, research experimentation, and multimodal application development.
 
 ## Key Highlights
 
-* **Unredacted MAX Training**: Fine-tuned to significantly reduce refusal patterns and improve instruction adherence across diverse prompts.
-* **4B Parameter Architecture**: Built on top of Qwen3-VL-4B-Instruct, balancing strong reasoning performance with more efficient hardware requirements compared to larger 8B variants.
-* **Unrestricted Multimodal Reasoning**: Designed for deep analysis of artistic, forensic, technical, or abstract visual content without standard safety-driven refusals.
-* **High-Fidelity Captions**: Produces dense, descriptive outputs suitable for dataset generation, metadata enrichment, or accessibility use cases.
-* **Dynamic Resolution Support**: Retains Qwen3-VL’s ability to process varying image resolutions and aspect ratios effectively.
+* **Optimized Release Packaging**
+  Streamlined repository structure for smoother loading, inference, and deployment workflows.
+
+* **Modern Transformers Compatibility**
+  Updated for stable integration with recent Hugging Face Transformers ecosystems.
+
+* **4B Vision-Language Architecture**
+  Built on **Qwen3-VL-4B-Instruct**, balancing multimodal capability with relatively low compute requirements.
+
+* **Stable Multimodal Inference**
+  Designed for consistent performance across image-text reasoning tasks.
+
+* **Efficient Caption Generation**
+  Produces structured and descriptive outputs suitable for annotation pipelines and dataset creation.
+
+* **Dynamic Resolution Support**
+  Retains native support for varying image resolutions and aspect ratios.
+
+---
+
+## Base Model Signatures:
+
+This model has been re-sharded and optimized for the latest Transformers version from the base model:
+https://huggingface.co/huihui-ai/Qwen3-VL-4B-Instruct-abliterated
+
+---
 
 ## Quick Start with Transformers
 
-```python
+```python id="vl4b_code"
 from transformers import Qwen3VLForConditionalGeneration, AutoProcessor
 from qwen_vl_utils import process_vision_info
 import torch
 
-# Load the 4B Unredacted MAX model
 model = Qwen3VLForConditionalGeneration.from_pretrained(
     "prithivMLmods/Qwen3-VL-4B-Instruct-Unredacted-MAX",
     torch_dtype="auto",
@@ -59,7 +79,7 @@ messages = [
                 "type": "image",
                 "image": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg",
             },
-            {"type": "text", "text": "Provide a detailed caption and reasoning for this image."},
+            {"type": "text", "text": "Provide a detailed caption for this image."},
         ],
     }
 ]
@@ -80,12 +100,8 @@ inputs = processor(
 
 generated_ids = model.generate(**inputs, max_new_tokens=256)
 
-generated_ids_trimmed = [
-    out_ids[len(in_ids):] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
-]
-
 output_text = processor.batch_decode(
-    generated_ids_trimmed,
+    [out[len(inp):] for inp, out in zip(inputs.input_ids, generated_ids)],
     skip_special_tokens=True,
     clean_up_tokenization_spaces=False
 )
@@ -93,27 +109,23 @@ output_text = processor.batch_decode(
 print(output_text)
 ```
 
-## Intended Use
-
-* **Advanced Red-Teaming**: Evaluating multimodal robustness and probing behavioral edge cases.
-* **Complex Data Archiving**: Generating detailed captions for medical, artistic, historical, or research datasets.
-* **Refusal Mechanism Research**: Studying behavioral shifts in vision-language models after abliterated fine-tuning.
-* **Creative Storytelling**: Producing detailed visual descriptions for narrative and world-building projects.
+---
 
-## Limitations & Risks
+## Intended Use
 
-> **Critical Note**: This model is designed to minimize built-in refusal mechanisms.
+* Multimodal AI research and evaluation
+* Vision-language dataset generation and annotation
+* Lightweight deployment on consumer GPUs
+* Prototyping multimodal applications and tools
+* Robustness testing and model benchmarking
 
-* **Sensitive Content Exposure**: The model may generate explicit or controversial descriptions if prompted accordingly.
-* **User Responsibility**: Generated outputs must be handled responsibly and used within ethical and legal boundaries.
-* **Hardware Requirements**: While lighter than 8B models, the 4B architecture still requires adequate VRAM for high-resolution image processing and longer generations.
+---
 
-## Acknowledgements
+## Limitations & Risks
 
-I would like to thank the works of the following:
+> **Important Note**: This model inherits limitations from its base architecture.
 
-* Uncensor any LLM with abliteration – [Maxime Labonne](https://huggingface.co/mlabonne)
-* Using FP8 and FP4 with Transformer Engine – [docs.nvidia](https://docs.nvidia.com/deeplearning/transformer-engine/user-guide/examples/fp8_primer.html)
-* Remove Refusals with Transformers – [Sumandora](https://github.com/Sumandora/remove-refusals-with-transformers)
-* LLM Compressor – [vllm-project](https://github.com/vllm-project/llm-compressor) 
-* FP8 Floating-Point 8: An Introduction to Efficient, Lower-Precision AI Training – [nvidia](https://developer.nvidia.com/blog/floating-point-8-an-introduction-to-efficient-lower-precision-ai-training/)
+* Output quality depends on image clarity and prompt design
+* May produce incomplete or inaccurate interpretations in complex scenarios
+* Requires sufficient VRAM for stable inference
+* Performance varies with decoding strategy and runtime optimizations