| --- |
| language: |
| - en |
| license: apache-2.0 |
| base_model: TurkishCodeMan/Qwen2.5-3B-Instruct-grpo-gmail |
| quantized_by: TurkishCodeMan |
| tags: |
| - gguf |
| - quantized |
| - tool-calling |
| - llama.cpp |
| - 4-bit |
| --- |
| |
| # Qwen2.5-3B-Instruct GRPO Gmail (Q4_K_M GGUF) |
|
|
| **Quantized version** of [TurkishCodeMan/Qwen2.5-3B-Instruct-grpo-gmail](https://huggingface.co/TurkishCodeMan/Qwen2.5-3B-Instruct-grpo-gmail). |
|
|
| ## 📥 Download & Run |
|
|
| \`\`\`bash |
| # Download (recommended - 3.5 GB) |
| huggingface-cli download TurkishCodeMan/Qwen2.5-3B-Instruct-grpo-gmail-GGUF Qwen2.5-3B-Instruct-grpo-gmail-Q4_K_M.gguf |
|
|
| # Run with GPU |
| ./llama-server -m Qwen2.5-3B-Instruct-grpo-gmail-Q4_K_M.gguf --port 8080 -ngl 99 |
|
|
| # Run on CPU |
| ./llama-server -m Qwen2.5-3B-Instruct-grpo-gmail-Q4_K_M.gguf --port 8080 |
| \`\`\` |
|
|
| ## ⚙️ Quantization Info |
|
|
| - **Method**: Q4_K_M (4-bit with K-means) |
| - **Size**: ~2.3 GB (vs 6.7 GB F16) |
| - **Quality**: 95%+ of F16 performance |
| - **Speed**: 3-4x faster inference |
|
|
| ## 🔗 Related Models |
|
|
| - **Full precision (F16)**: [TurkishCodeMan/Qwen2.5-3B-Instruct-grpo-gmail](https://huggingface.co/TurkishCodeMan/Qwen2.5-3B-Instruct-grpo-gmail) |
| - **Base model**: [unsloth/Qwen2.5-3B-Instruct](https://huggingface.co/unsloth/Qwen2.5-3B-Instruct) |
|
|
| ## 🎯 Tool Calling Example |
|
|
| \`\`\`python |
| import requests |
|
|
| response = requests.post("http://localhost:8080/v1/chat/completions", json={ |
| "messages": [ |
| {"role": "system", "content": "You are a tool-calling assistant."}, |
| {"role": "user", "content": "Send email to test@gmail.com about meeting tomorrow"} |
| ], |
| "temperature": 0.0, |
| "max_tokens": 512 |
| }) |
| |
| print(response.json()['choices'][0]['message']['content']) |
| # Output: {"tool_calls": [{"function": "send_email", "arguments": {"to": ["test@gmail.com"], "subject": "Meeting Tomorrow", "body": "..."}}]} |
| \`\`\` |
|
|
| ## 📊 Training |
|
|
| - **SFT**: 300 steps on 57 Gmail examples |
| - **GRPO**: 300 steps reinforcement learning for tool calling accuracy |
| - **Final loss**: 0.50 (excellent convergence) |
|
|
| ## 🛠️ Supported Tools |
|
|
| \`send_email\`, \`draft_email\`, \`read_email\`, \`search_emails\`, \`delete_email\`, \`modify_email\`, \`batch_modify_emails\`, \`batch_delete_emails\`, \`list_email_labels\`, \`create_label\`, \`update_label\`, \`delete_label\`, \`get_or_create_label\`, \`create_filter\`, \`list_filters\`, \`get_filter\`, \`delete_filter\`, \`create_filter_from_template\`, \`download_attachment\` |
|
|