Datasets:
Questions, results and reproduction notes
#1
by CompilingThings - opened
Welcome. This repo holds a fine-tuned Qwen2.5-Coder-14B for MQL5 (MetaTrader 5) code generation,
evaluated on a held-out set of 184 tasks scored by compilation against the real MetaEditor
toolchain, with the frozen prompts, per-item results and checksums included so the comparison can
be reproduced.
Use this thread for questions about the setup, reproduction results on your own hardware, or
anything in the evaluation that looks wrong. Corrections are welcome and will be applied in place
with a note, not by re-posting. Please open a separate discussion for feature requests or
unrelated MQL5 questions.