Update README.md

README.md

@@ -3,21 +3,169 @@ tags:
 - gguf
 - llama.cpp
 - unsloth
+- mistral
+- python
+base_model:
+- mistralai/Mistral-7B-Instruct-v0.3
+---
+
+# mistral-7b-python-gguf
+
+Conversational Python fine-tune of Mistral 7B exported to GGUF format for local inference.
+
+- Base model: Mistral 7B
+- Fine-tuning framework: Unsloth
+- Format: GGUF
+- Author: AntoineChatry
 
 ---
 
-# mistralv1-gguf : GGUF
+# ⚠️ Disclaimer
+
+This is an **early experimental fine-tune**.
+
+It is **not production-ready**, not fully aligned, and not optimized for reliability or long-form reasoning.  
+This project was created primarily for learning and experimentation.
+
+Please do not expect state-of-the-art coding performance.
+
+---
+
+# Model Overview
+
+This model is a conversational fine-tune of Mistral 7B trained primarily on:
+
+- ShareGPT-style conversations
+- Python-focused discussions
+- Coding Q&A format
+
+The objective was to:
+
+- Experiment with fine-tuning
+- Build a conversational Python model
+- Export to GGUF for llama.cpp compatibility
+- Test local inference workflows
+
+No RLHF or advanced alignment was applied beyond the base model.
+
+---
 
-This model was finetuned and converted to GGUF format using [Unsloth](https://github.com/unslothai/unsloth).
+# Known Limitations
+
+## Repetition Issues
+
+- Frequently repeats phrases like:
+  > "Here's the code:"
+- Can loop or restate similar sentences
+- Overuses patterns learned from dataset formatting
+
+## Weak Long-Form Explanations
+
+- Struggles with multi-paragraph structured reasoning
+- May repeat itself when asked for detailed explanations
+- Limited depth on conceptual explanations
+
+## Instruction Following
+
+- Not fully aligned
+- May ignore strict formatting constraints
+- Tends to prioritize generating code over detailed explanations
+
+## Dataset Bias
+
+- Strong ShareGPT conversational tone
+- Python-heavy bias
+- Some templated response structure
+
+---
+
+# What Works Reasonably Well
+
+- Short Python snippets
+- Basic debugging help
+- Simple function generation
+- Conversational coding prompts
+
+Best performance is observed when:
+
+- Prompts are clear and direct
+- Expected output is short
+- Tasks are code-focused
+
+---
 
-**Example usage**:
-- For text only LLMs:    `llama-cli -hf AntoineChatry/mistralv1-gguf --jinja`
-- For multimodal models: `llama-mtmd-cli -hf AntoineChatry/mistralv1-gguf --jinja`
+# Training Details
 
+- Base: Mistral 7B
+- Dataset format: ShareGPT-style conversational dataset (Python-oriented)
+- Fine-tuned using Unsloth notebooks
+- Converted to GGUF for llama.cpp compatibility
+- Quantized version included (Q4_K_M)
+
+No additional safety tuning or post-training optimization was applied.
+
+---
+
+# Example Usage
+
+This model was finetuned and converted to GGUF format using Unsloth.
+
+## llama.cpp
+
+For text-only LLMs:
+
+```bash
+llama-cli -hf AntoineChatry/mistral-7b-python-gguf --jinja
+```
+
+For multimodal models:
+
+```bash
+llama-mtmd-cli -hf AntoineChatry/mistral-7b-python-gguf --jinja
+```
+
+---
 ## Available Model files:
 - `mistral-7b-instruct-v0.3.Q4_K_M.gguf`
+---
+# Ollama
 
-## Ollama
 An Ollama Modelfile is included for easy deployment.
-This was trained 2x faster with [Unsloth](https://github.com/unslothai/unsloth)
-[<img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png" width="200"/>](https://github.com/unslothai/unsloth)
+
+Example:
+
+```bash
+ollama create mistral-python -f Modelfile
+ollama run mistral-python
+```
+
+---
+
+# Why This Model Is Public
+
+This model represents a learning milestone.
+
+Sharing imperfect models helps:
+
+- Document fine-tuning progress
+- Enable experimentation
+- Collect feedback
+- Iterate toward better versions
+
+This is not a finished product.
+
+---
+
+# Unsloth
+
+This model was trained 2x faster using Unsloth.
+
+https://github.com/unslothai/unsloth
+
+<img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png" width="200"/>
+
+---
+
+# License
+
+Please refer to the original Mistral 7B license from Mistral AI.