Fine-Tune Cooperativism
From the AI Research Compass wiki
Fine-Tune Cooperativism: Cheap, parameter-efficient fine-tuning lets every community adapt shared base models to its own needs, cooperatively and without frontier budgets.
Beliefs
Fine-Tune Cooperativism believes the most useful models are not the biggest ones but the ones adapted to a particular language, domain, style or task. Thanks to low-rank adapters and quantized training, that adaptation no longer requires a cluster: a single GPU and a curated dataset of modest size can specialize a strong open base model. The cooperative ideal is a commons of small, shareable adapters built by many groups on a few shared bases, each serving its own members while others can reuse, merge and improve them.
It sits near the Language edge, since its home turf is language-model adaptation, and close to the middle of the vertical axis: it builds on one shared base model, but puts capability into a distributed layer of adapters and datasets around it. Its paradigm level is exploratory: less a new theory of intelligence than a practical economics of adaptation that departs from the one-big-model mainstream.
Arguments
- Low-rank adaptation matches full fine-tuning on many tasks while training only a small fraction of parameters, making adapters cheap to train, store and share.
- Quantized low-rank fine-tuning made it practical to adapt large models on a single GPU.
- Adapters are modular: one base model can serve many specialized variants, swapped at inference time.
Critiques
- Fine-tuning mostly reshapes behavior and format; it adds far less new knowledge or reasoning ability than practitioners sometimes hope.
- Small fine-tunes can erode a base model’s general abilities or safety behavior in ways that are hard to detect.
- Context-engineering advocates argue that retrieval and good prompting often achieve the same specialization with no training at all.
Representative works
- LoRA: Low-Rank Adaptation of Large Language Models2021
- QLoRA: Efficient Finetuning of Quantized LLMs2023
- Parameter-Efficient Transfer Learning for NLP2019
Relationships
Friends
Frenemies
How to get it
Land on this cell in the full or pro test: x ≈ −0.70, y ≈ −0.47, paradigm ≈ +0.28.