BEE-spoke-data/beecoder-220M-python

This is BEE-spoke-data/smol_llama-220M-GQA fine-tuned for code generation on:

• filtered version of stack-smol-XL
• deduped version of 'algebraic stack' from proof-pile-2
• cleaned and deduped pypi (last dataset)

This model (and the base model) were both trained using ctx length 2048.

examples

Example script for inference testing: here

It has its limitations at 220M, but seems decent for single-line or docstring generation, and/or being used for speculative decoding for such purposes.

The screenshot is on CPU on a laptop.

---