Upload README.md with huggingface_hub

README.md

@@ -4,71 +4,93 @@ base_model: Qwen/Qwen3-8B
 tags:
   - crewai
   - code-generation
-  - fine-tuned
+  - multi-agent
+  - qwen3
   - gguf
-  - ollama
-language:
-  - en
-pipeline_tag: text-generation
+  - unsloth
+  - thinking
 ---
 
-# CrewAI Qwen3-8B GGUF
+# CrewAI Code Generation - Qwen3 8B (GGUF)
 
-Fine-tuned Qwen3-8B model for generating CrewAI multi-agent code.
+Fine-tuned Qwen3-8B models for generating CrewAI multi-agent code from natural language descriptions.
 
-## Model Description
+## Models Available
 
-This model was fine-tuned on 2,500 high-quality CrewAI code examples to generate complete CrewAI crews including:
-- Agent definitions (role, goal, backstory, tools)
-- Task definitions with context dependencies
-- Crew instantiation with process type
-- Kickoff code with inputs
+### V2 Thinking Models (Recommended)
+| File | Size | Description |
+|------|------|-------------|
+| crewai-qwen3-8b-thinking-v2-q4_k_m.gguf | 4.7 GB | 4-bit quantized, best balance |
+| crewai-qwen3-8b-thinking-v2-q8_0.gguf | 8.2 GB | 8-bit quantized, higher quality |
+
+**V2 Improvements:**
+- Native Qwen3 chat template (ChatML format)
+- Higher LoRA rank (r=32 vs r=16)
+- Thinking mode with reasoning tags
+- Better structured outputs
+
+### V1 Models (Legacy)
+| File | Size | Description |
+|------|------|-------------|
+| qwen3-8b.Q4_K_M.gguf | ~4.7 GB | 4-bit quantized |
+| qwen3-8b.Q5_K_M.gguf | ~5.5 GB | 5-bit quantized |
+| qwen3-8b.Q8_0.gguf | ~8.2 GB | 8-bit quantized |
 
 ## Usage with Ollama
 
-1. Download the GGUF file (Q4_K_M recommended for most users)
+1. Download the GGUF file
 2. Create a Modelfile:
 
 ```
-FROM ./qwen3-8b.Q4_K_M.gguf
-
-TEMPLATE """Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.
-
-### Instruction:
-{{ .System }}
+FROM ./crewai-qwen3-8b-thinking-v2-q4_k_m.gguf
+
+TEMPLATE """{{- if .System }}
+<|im_start|>system
+{{ .System }}<|im_end|>
+{{ end }}
+<|im_start|>user
+{{ .Prompt }}<|im_end|>
+<|im_start|>assistant
+"""
 
-### Input:
-{{ .Prompt }}
+SYSTEM """You are a CrewAI code generation expert. When given a task description and required inputs, generate complete, working CrewAI Python code that includes:
+- All necessary imports (crewai, crewai_tools)
+- Agent definitions with roles, goals, backstories, and tools
+- Task definitions with proper context dependencies
+- Crew instantiation with appropriate process type
+- Kickoff code with the provided inputs
 
-### Response:
-"""
+Think through the problem step by step before generating code."""
 
 PARAMETER temperature 0.7
 PARAMETER top_p 0.9
-PARAMETER repeat_penalty 1.2
-PARAMETER stop "### Instruction:"
-PARAMETER stop "### Input:"
 PARAMETER num_ctx 8192
+PARAMETER stop "<|im_end|>"
+PARAMETER stop "<|im_start|>"
 ```
 
-3. Create the model: `ollama create crewai-qwen3-8b -f Modelfile`
+3. Create and run:
+```bash
+ollama create crewai-qwen3-8b -f Modelfile
+ollama run crewai-qwen3-8b
+```
 
-## Quantization Options
+## Example Prompt
 
-| File | Size | Description |
-|------|------|-------------|
-| Q4_K_M | ~4.7 GB | Good balance of quality/size (recommended) |
-| Q5_K_M | ~5.5 GB | Better quality, larger size |
-| Q8_0 | ~8.2 GB | Best quality, largest size |
+```
+Create a CrewAI crew for analyzing competitor websites.
+Required inputs:
+- competitor_urls: list of URLs to analyze
+- analysis_focus: what aspects to focus on
+```
 
 ## Training Details
 
-- Base model: Qwen/Qwen3-8B
-- Training framework: Unsloth
-- Dataset: 2,500 CrewAI code examples in Alpaca format
-- Epochs: 3
-- LoRA rank: 16
+- **Base Model**: Qwen/Qwen3-8B
+- **Fine-tuning**: Unsloth + QLoRA (r=32, alpha=32)
+- **Dataset**: 2,500 CrewAI code examples
+- **Training**: 3 epochs on RTX 4090
 
 ## License
 
-Apache 2.0
+Apache 2.0 (same as base Qwen3 model)