metadata
license: other
license_name: research-only-archive
base_model: Qwen/Qwen2.5-7B-Instruct
language:
- id
- en
tags:
- migancore
- archived
- research
- orpo
- lora
MiganCore 7B "soul" v0.2 — archived early experiment (May 2026)
Status: archived, superseded, not maintained. Published for research transparency only. This is an early iteration of MiganCore's Qwen2.5-7B line. The project later moved to a 4B base and closed on 28 September 2026.
What it is
- Base: Qwen/Qwen2.5-7B-Instruct (Apache-2.0).
- Method: ORPO / APO-zero with a LoRA adapter (r = 16, learning rate 5e-7, 1 epoch).
Training data
613 preference pairs. About 97 % of the preferred answers were written by a commercial LLM (identity, tool-use and code-correctness pairs, with hard-coded rejected answers); the rest carry over from v0.1.
Because training data came from the outputs of third-party commercial LLM services, check the terms of those services before any use beyond research.
Evaluation
No pre-registered evaluation exists for this version.
More
- Lineage of all 41 MiganCore variants and why this line was archived: github.com/fahmiwol/migancore
- Method and closing report: github.com/fahmiwol/migancore-research-method
Author: Fahmi Ghani (fahmiwol@gmail.com).