srivarenya/MoM-python-slm-grpo
The srivarenya/MoM-python-slm-grpo is a 1.5 billion parameter Python code generation model, warm-started from Qwen2.5-Coder-1.5B-Instruct and fine-tuned using Group Relative Policy Optimization (GRPO). This model is specifically designed as a spec-driven code-generation node within a Mixture-of-Models (MoM) mesh, excelling at generating Python functions from specifications. It demonstrates significant improvements on benchmarks like MBPP (+2.9) and domain problem_solving (+5.4) compared to its SFT predecessor, making it highly effective for tasks requiring correct function generation from detailed requirements.
Loading preview...
MoM-Python-SLM-GRPO: Spec-Driven Python Code Generation
This model, developed by srivarenya, is a 1.5 billion parameter Python code generation model, serving as the spec-driven node within a Mixture-of-Models (MoM) mesh. It is a GRPO/RLVR-tuned successor to srivarenya/MoM-python-slm, warm-started from a DoRA r=64 SFT of Qwen2.5-Coder-1.5B-Instruct.
Key Capabilities & Differentiators
- Optimized for Spec-Driven Code: Fine-tuned using Group Relative Policy Optimization (GRPO) with a reward function heavily weighted towards execution success (0.8 execution + 0.1 format + 0.1 LLM-judge).
- Strong Performance on MBPP: Achieves 72.5 on MBPP (greedy pass@1), a +2.9 improvement over its SFT sibling, and +5.4 on the
problem_solvingdomain. - Specialization Trade-off: While excelling at generating correct functions from specifications, it shows a slight decrease on HumanEval completion tasks (−3.0), reflecting its specialized reinforcement learning focus.
- Context Length: Supports a context length of 32768 tokens.
When to Use This Model
- Spec-Driven Code Generation: Ideal for tasks where the primary goal is to generate a correct Python function based on a given specification or problem description.
- MoM Architectures: Designed to function as a specialized node within a larger Mixture-of-Models system for enhanced code generation.
For HumanEval-completion heavy use cases, the SFT sibling might be more suitable.