my-ai-stack/Stack-4.0-Qwen-3B-Merged
my-ai-stack/Stack-4.0-Qwen-3B-Merged is a 3.1 billion parameter instruction-tuned coding model, fully merged from Qwen2.5-Coder-3B-Instruct. Developed by my-ai-stack, this model incorporates 55,000 agentic tool-use conversations, making it specialized for generating code and handling agentic tasks. It operates as a standalone model without requiring adapters and supports a 32768 token context length.
Loading preview...
Stack 4.0 Omni-Nexus — Merged
This model, my-ai-stack/Stack-4.0-Qwen-3B-Merged, is a 3-billion parameter instruction-tuned coding model. It is fully merged from Qwen2.5-Coder-3B-Instruct and includes 55,000 agentic tool-use conversations, enabling it to generate tool calls for agentic workflows.
Key Capabilities
- Coding Focus: Specifically designed and fine-tuned for code generation tasks.
- Agentic Tool-Use: Incorporates extensive training on agentic tool-use conversations.
- Standalone Deployment: Provided as a fully merged model, eliminating the need for LoRA adapters or base model dependencies, allowing direct deployment.
- Performance: Achieves 74.0% on HellaSwag (acc_norm) and 52.0% on ARC-Challenge (acc_norm) based on 50-sample evaluations, with internal coding samples producing valid Python.
Training Details
The model was trained using a QLoRA method, then merged, with 7.3 million trainable parameters (0.24% of 3.1B). Training involved 1,000 steps over approximately 10 hours on GCP Tesla V100 16GB hardware, with a final training loss of 0.1411.
Limitations
- Size: As a 3B model, it may be less capable for complex, multi-step reasoning compared to larger models.
- Language Optimization: Primarily optimized for English; performance in other languages may vary.
- Tool Execution: While it generates tool calls, actual execution requires an external agent loop in the application.