Compresser et fine-tuner un modèle

The family of techniques that either shrinks a model or adapts it to your need. Quantization: weights are stored at lower precision (16-bit down to 4-bit), same model, just smaller and faster. Distillation: a small student model learns to imitate a large teacher model, you get a different, lighter model. LoRA: you fine-tune by training only a small set of added weights, not all of them, so adapting a huge model becomes cheap; QLoRA does the same after quantizing the base to 4-bit, to fit on a single consumer GPU. SFT versus RLHF: training on example answers to copy (imitation) versus training on what humans actually preferred (preference optimization). Prefill versus decode: reading the whole prompt in one parallel pass versus generating the answer one token at a time.

Strengths

Limitations

Best for

Official site

View on Coeurdar