Kyle O'Brien's picture

Kyle O'Brien PRO

Kyle1668

·

https://kyleobrien.io

Kyle1668

AI & ML interests

pretraining, alignment, open-source

Recent Activity

updated a dataset 1 day ago

geodesic-research/pa-warm-start-1B-sft-mix

new activity 1 day ago

geodesic-research/pa-warm-start-1B-sft-mix:Migrate tool_calls/tools from JSON strings to structured columns OpenAI-convention hybrid: tool_calls is list<struct{id,type,function{name,arguments}}>, tools is list<struct{type,function{name,description,parameters}}>; arguments/parameters remain JSON-encoded strings (Arrow-clean across heterogeneous tools). JSON-string tool_calls char-iterate in Jinja chat templates, rendering one empty <tool_call><function=></function></tool_call> block per character — 4-5x length blowup and a deterministic training NaN. Renders validated byte-identical to the parsed old rows on every config.

new activity 1 day ago

geodesic-research/pa-warm-start-1B-sft-mix:Restore `default` config in the configs: mapping The explicit top-level `configs:` section (added by the per-config pushes) shadows the implicit default config, so load_dataset(repo, 'default') fails with "BuilderConfig 'default' not found" even though data/ holds the blended mix. This adds the data_files mapping for it (data/train-*, 265,048 rows).

View all activity

Organizations

Kyle1668 's papers 5

arxiv:2508.06601

arxiv:2407.06483

arxiv:2406.17746

arxiv:2402.08225

arxiv:2304.01373