Local model research
AVA
How much can a small model learn?
A local AI project shaped by a laptop GPU: from scratch models to a released adapter, with the experiment history kept alongside the result.
- State
- v2 released · v3 in research
- Working with
- Python / PyTorch / QLoRA / llama.cpp
01 The constraint
AVA explores useful local AI within a 4 GB VRAM budget. The early work trained small models from scratch, then added retrieval and memory. Those experiments exposed a knowledge ceiling. The next step was to adapt a pretrained model while keeping training and evaluation practical on the same laptop.
02 The decisions
Change the starting point.
Scratch training made the tokenizer, retrieval and memory mechanisms inspectable, but the bare model did not establish useful ARC reasoning. QLoRA on Qwen3.5-2B offered a different trade-off: retain pretrained knowledge and train a small adapter rather than every weight.
Treat a completed run as a question.
The first fine-tuning attempt ran within the hardware budget but did not improve ARC over the base model in the recorded evaluation. The v2 work expanded the curated training material and improved the training path instead of treating a successful run as proof of better reasoning.
Release the evidence with the artifact.
The repository includes a training recipe, a benchmark report, evaluation files and an export path. The adapter and GGUF releases make the result usable locally; retained failed and paused experiments explain why the project changed direction.
- 01Curated examples
- 02QLoRA adapter
- 03Local export
- 04Evaluation
03 What exists
The repository documents a released AVA v2 adapter and quantized local models, plus a full evaluation report with scoring methodology and confidence intervals. The useful output is both a runnable artifact and a record of how a constrained training system developed.
04 Still open
The v2 results show weaker maths, narrative commonsense and tool use. Training used short sequences, and the multiple-choice scoring protocol differs from common leaderboard protocols. Cross-model tables are context rather than controlled comparisons. AVA v3 is a separate research track; its plans do not establish a released capability.
Follow the evidence
Based on project code and documentation reviewed in October 2026. Evaluation results are reported by the projects; they were not independently rerun for this portfolio.
- Experiment history
Scratch models, the first adapter, the v2 release and the paused branches.
- Results and caveats
Repository-reported evaluations, scoring protocol and known weaknesses.
- v2 training implementation
The actual adapter configuration and training entry point.
- Released adapter
The model artifact linked by the project documentation.