David Thrower

## From #260

integrate-suggested-solution-to-streaming-llm-phase-I-b-training-data

1

# Add prototype for Phase I-B training with a tf.data.Dataset (from generator) ## From #255 ## TLDR Stream the operation performed in `prepare_data()` in https://github.com/david-thrower/cerebros-core-algorithm-alpha/blob/255-copy-of-branch-254-updated-hpo-script-for-cicd-scale-testing/generative-proof-of-concept-CPU-preprocessing-in-memory.py ... with numerous arbitrary constraints...

# Bug file name mismatch if there is a capital letter in the project name If I instantiate cerebros.simple_random_search.SimpleRandomSearch(... project_name="anythingWithCapitalLetters"), the will be an error when it looks for the...

Currently, jit_compile is set based on whether a base model is being used, because historically, most base models were not compatible with the current generation CPU's XLA capabilities. This logic...

# from #254 : A CICD scale copy of the HPO script for further study.

# Tweaks and refactors to ## **From #253** ## TO DO: - [x] ### Add to text generation samples: - temperature=0.75, top_k=75, top_p=0.98, presence_penalty=1.4, frequency_penalty = 1.4 - temperature=0.7, top_k=75,...

# Run an HPO study training NotGPT on the 681 sample World English Bible Genesis 40 token context length data set. ## From #252

# From #240

# Test that this will run at the small scale to lint the larger scale container. ```Dockerfile FROM tensorflow/tensorflow:2.19.0 RUN apt update -y RUN apt upgrade -y RUN apt install...

2025-09-27--new-hpo-study-with-sampling-plus-penalties

1

# Reproduce #243 with penalties and perhaps a more tuned hyperparam range considering the results of the last run. ## From #240 (latest updates to API including penalties for repetition)