by stas
Open source · 309k downloads · 41 likes
The *tiny random llama 2* model is a lightweight and randomized version of the well-known Llama 2, derived from the original 7-billion-parameter model. Designed primarily for functional testing, it does not produce high-quality outputs, as its weights are randomly generated and its vocabulary is limited to 3,000 tokens. It enables quick validation of a pipeline or application integration without requiring significant computational resources. Its purpose is solely to verify that interactions with a language model are functioning correctly, without claiming to generate coherent or relevant responses. This model stands out for its simplicity and efficiency, making it ideal for preliminary testing or basic demonstrations.
This is a tiny random Llama model derived from "meta-llama/Llama-2-7b-hf".
See make_tiny_model.py for how this was done.
This is useful for functional testing (not quality generation, since its weights are random and the tokenizer has been shrunk to 3k items)