Blair Johnson
@blrjohnson.bsky.social
ML Ph.D. Student @ Georgia Tech,
Research Engineer @ GTRI,
I work on symbolic pattern discovery in graphs.
Research Engineer @ GTRI,
I work on symbolic pattern discovery in graphs.
Did a quick experiment similar to this last year and observed that you can definitely bias the distribution of response lengths to a given prompt. github.com/Blair-Johnso...
GitHub - Blair-Johnson/tokens-to-completion: Experiment related to controlling LLM output length
Experiment related to controlling LLM output length - Blair-Johnson/tokens-to-completion
github.com
November 25, 2024 at 12:54 AM
Did a quick experiment similar to this last year and observed that you can definitely bias the distribution of response lengths to a given prompt. github.com/Blair-Johnso...
Do you know of a good source that documents some of these tricks in detail? The Jax docs are very good IMO, but I would be interested in some lessons learned from a user’s perspective.
November 23, 2024 at 7:06 PM
Do you know of a good source that documents some of these tricks in detail? The Jax docs are very good IMO, but I would be interested in some lessons learned from a user’s perspective.