didula-wso2/gemma4_1-0-8_sft_16bit_vllm
The didula-wso2/gemma4_1-0-8_sft_16bit_vllm model is a 7.9 billion parameter language model, fine-tuned from unsloth/gemma-4-E4B-it using a Ballerina SFT process. This model specializes in code generation, demonstrating strong performance on HumanEval-style benchmarks for both Ballerina and Python. Its primary differentiation lies in its specific optimization for programming tasks, making it suitable for code-related applications.
Loading preview...
Model Overview
The didula-wso2/gemma4_1-0-8_sft_16bit_vllm is a 7.9 billion parameter language model, specifically a 16-bit merged version of a Ballerina Supervised Fine-Tuning (SFT) of unsloth/gemma-4-E4B-it. The fine-tuning process involved LoRA (r16/alpha32) applied to all attention and MLP projections, trained for one epoch over 9,182 samples with rules-free system prompts.
Key Capabilities
- Code Generation: The model is explicitly fine-tuned for programming tasks, particularly with Ballerina and Python.
- Performance on HumanEval: It achieves 90/159 on a Ballerina HumanEval-style benchmark and a 94.3% pass@1 rate on Python HumanEval (3 samples, t=0.8), indicating strong coding abilities.
- Fine-tuning Details: The SFT used a learning rate of 1e-4 with a cosine schedule and a seed of 3407. The adapter was merged into the bf16 base on the language-model modules.
Good For
- Code-centric applications: Ideal for tasks requiring code generation, completion, or understanding, especially in Ballerina and Python.
- Developers working with Ballerina: The specific Ballerina SFT suggests a tailored performance for this language.
- Benchmarking code models: Its reported HumanEval scores provide a clear indication of its capabilities in competitive coding scenarios.