Tiny Pickle v3 Coder is a coding-focused LoRA adaptation of
Qwen/Qwen3-Coder-30B-A3B-Instruct.
It was trained for ten hours on examples from OpenCodeInstruct for which
the recorded average test score was perfect and every recorded test
passed.
Training
Optimizer steps: 9,920
Approximate packed sequences: 39,680
Training context: 4,096 tokens
LoRA rank: 64
Learning rate: 2e-5
Hardware: one NVIDIA H200
Training duration: ten hours
Intended use
Code generation, debugging, implementation planning, code review,
algorithmic problem solving and test generation.
Limitations
This is an experimental specialization. It has not yet been demonstrated
to outperform the unmodified base model on independent benchmarks.
Generated code may be incorrect, insecure or incomplete and must be
reviewed and tested.
