Replies: 4 comments 1 reply
-
Good research 👍 . I've also found that the alpaca-lora models get easily stuck in a loop. However, using the palpaca-7B (non-lora, q4_0) model, it hasn't gotten stuck in a loop, not once, when running in instruct mode with default settings and |
Beta Was this translation helpful? Give feedback.
-
For reference the same with palpaca-7B q4_0
Your poem seems a lil better tbh 😄 Neither are particularly amazing 😆 |
Beta Was this translation helpful? Give feedback.
-
The output definitely changed with
|
Beta Was this translation helpful? Give feedback.
-
@blackle I just pushed a draft version of a new repeat penalization heuristic that might help with sampling at low temperatures #331 (comment) |
Beta Was this translation helpful? Give feedback.
-
If you use
--top_p 0.0 --top_k 1
then the sampler will always pick the most likely next token. For the 7b alpaca model I have, this works extremely well and rarely gets caught in loops.As an example, here's a prompt:
And the output:
Although the network still hallucinates sometimes, I find that greedy sampling makes it less common and also it sticks to the prompt better. This is especially useful if you have a specific task, like if you want it to extract information from some unstructured input.
Beta Was this translation helpful? Give feedback.
All reactions