PyBytecode
Copyright 2026 Blazing Customs

This product includes software developed from Qwen2.5-Coder-1.5B-Instruct
(https://huggingface.co/Qwen/Qwen2.5-Coder-1.5B-Instruct), Copyright Alibaba Cloud,
licensed under the Apache License, Version 2.0.

The upstream repository ships a LICENSE file and no NOTICE file (checked 2026-08-04 against
the Hugging Face API), so there is no upstream NOTICE text to append here.

CHANGES MADE TO THE LICENSED WORK, as required by Apache-2.0 section 4(b):

  The base model was fine-tuned with LoRA (rank 16, alpha 32, applied to q_proj, k_proj,
  v_proj, o_proj, gate_proj, up_proj and down_proj; 1 epoch; learning rate 2e-4) on 48,196
  pairs of Python 3.12 bytecode disassembly and the source that produced it. The resulting
  adapter was merged into the base weights. No architecture, vocabulary or tokenizer change
  was made.

The training corpus is not distributed. Its per-row attribution was not retained, so it cannot
be redistributed without stripping required notices from its upstream authors. See
DATA-CARD-training-corpus.md. This constrains the corpus only; it does not encumber these
weights, which are trained parameters rather than a copy of any source text, and which were
derived from a corpus pre-filtered to seven permissive licences with all GPL, LGPL, AGPL, MPL
and EPL rows dropped before extraction.