MEGAN 1.0 — Coder 7B LoRA
This repository contains the persistent LoRA adapter artifacts for the MEGAN 1.0 coding-model project.
Base model
Qwen/Qwen2.5-Coder-7B-Instruct
The base model is not duplicated in this repository. MEGAN loads the base model separately and applies the adapter stored here.
Training status
The initial checkpoint in this repository is a QLoRA pipeline smoke-test checkpoint.
It was created to verify:
- 4-bit NF4 model loading
- LoRA attachment
- forward pass
- backward pass
- optimizer step
- checkpoint saving
- checkpoint reloading
- inference after reload
- Hugging Face persistence
It is NOT a benchmark-quality MEGAN release and should not be interpreted as evidence of improved coding performance.
Hardware validation
Validated on a Google Colab Tesla T4 with approximately 14.56 GB VRAM.
Important
Future production MEGAN checkpoints should be trained on a properly sourced, provenance-tracked coding dataset with separate validation and held-out evaluation data.
The training project should preserve checkpoint metadata, evaluation results, dataset provenance, and reproducibility information.
License
The base model's license and terms remain applicable to use of the base model.
The adapter artifacts in this repository should be used consistently with the applicable base-model and dataset licenses.
- Downloads last month
- 18