Our great sponsors
-
InfluxDB
Power Real-Time Data Analytics at Scale. Get real-time insights from all types of time series data with InfluxDB. Ingest, query, and analyze billions of data points in real-time with unbounded cardinality.
You can use your own GPU to host and fine-tune LLMs with Refact self-hosting server.
It's possible to fine-tune all parameters (called "full fine-tune"), but recently PEFT methods became popular. PEFT stands for Parameter-Efficient Fine-Tuning. There are several methods available, the most popular so far is LoRA (2106.09685) that can train less than 1% of the original weights. LoRA has one important parameter -- tensor size, called lora_r. It defines how much information LoRA can add to the network. If your codebase is small, the fine-tuning process will see the same data over and over again, many times in a loop. We found that for a smaller codebase small LoRA tensors work best because it won't overfit as much -- the tensors just don't have the capacity to fit the limited training set exactly. As the codebase gets bigger, tensors should become bigger as well. We also unfreeze token embeddings at a certain codebase size. To pick all the parameters automatically, we have developed a heuristic that calculates a score based on the source files it sees. This score is then used to determine the appropriate LoRA size, number of finetuning steps, and other parameters. We have tested this heuristic on several beta test clients, small codebases of several files, and large codebases like the Linux kernel (consisting of about 50,000 useful source files). If the heuristic doesn't work for you for whatever reason, you can set all the parameters yourself.
Related posts
- DECT NR+: A technical dive into non-cellular 5G
- Training LLMs Taking Too Much Time? Technique you need to know to train it faster
- OpenAI employee: GPT-4.5 rumor was a hallucination
- Can a LoRa be used on models other than Stable Diffusion?
- Andreessen Horowitz Invests in Civitai, Which Profits from Nonconsensual AI Porn