Lifecycle Mentor LoRA pilot
Experimental Portuguese-language LoRA adapter for educational assistance with product development (PDLC) and software development (SDLC).
Base model and use
This repository contains only the MLX LoRA adapter, not the base model. Load it with mlx-community/Qwen2.5-7B-Instruct-4bit and a compatible mlx-lm version. The base model is Qwen/Qwen2.5-7B-Instruct (Apache-2.0). The adapter uses CC-BY-SA-4.0 to match the license declared for the fictional examples used in this pilot; review the source dataset and base model terms before redistribution or reuse.
Training data and limitations
The initial pilot was trained on six human-reviewed, fictional demonstration examples from the Lifecycle Mentor repository: four train, one validation, and one test example. This is far too small to support claims of general performance, safety, or factual reliability. The adapter is an engineering experiment and must be evaluated with a larger, independently reviewed dataset before production use.
Pilot run: 20 LoRA iterations; validation loss moved from 4.367 to 4.176. The single held-out example produced test loss 3.890 and perplexity 48.906. These values are included for reproducibility only and are not a meaningful benchmark.
No real student conversations or personal data are included. The examples are marked fictional and CC-BY-SA-4.0 in the source dataset.
Intended use
Educational prototyping and evaluation of Portuguese PDLC/SDLC guidance. The model is not a source of medical, legal, or other professional advice and must not fabricate user research, metrics, or evidence.
Reproducibility
- Adapter repository:
tadrianonet/lifecycle-mentor - Base model:
Qwen/Qwen2.5-7B-Instruct - Training framework: MLX-LM LoRA on Apple Silicon
- Configuration:
training/config.yamlin the source project - Training examples: 4; validation examples: 1; test examples: 1
Quantized