Spaces:
Running on CPU Upgrade
Running on CPU Upgrade
Commit History
12 days ago
-
Rework COLM spotlight ending and notes; clarify unmixed baselines in blog
Slides: - Replace the headline-result bar chart with the blog's opening training-curve chart (3.2x faster, 31% better) and move it to second to last. - Reorder the ending to takeaways, recipe, cost, headline result, thank you. - Drop DROP from the trade-off slide (seed noise of 2 points). - Add name and affiliation to the closing slide. Notes: - Trim the spoken script to about 8 minutes, write billions as B, and avoid unexplained HQ. - Move all Q&A material to the last slide, including seed-variance numbers from the 72-run grid. Blog: state that the synthetic baselines in the opening and FinePhrase comparison figures are trained without a web mix-in, and add the 50/50-mixed baseline scores. Co-authored-by: Cursor <cursoragent@cursor.com>bd09275 -
fix TReB QA mislabel and PIQA category; add model-size slide and Q&A notes to COLM spotlight
The treb_qa benchmark (lighteval/treb_table_retrieval) was labeled TriviaQA in the blog and charts; relabel it and cite TReB. The macro score counts PIQA under Reasoning, so the setup list and correlation chart now group it there. Fix the reversed color legend in the score-correlation caption. The COLM deck gets the trained-model-size slide, a training-compute estimate and benchmark cheat sheet in the notes, the 4:30 poster time, and DROP instead of TriviaQA on the trade-off slide. Co-authored-by: Cursor <cursoragent@cursor.com>c24632b
13 days ago
-
advertise poster session on COLM spotlight closing slide
Co-authored-by: Cursor <cursoragent@cursor.com>c934bdf
18 days ago
-
add COLM 2026 spotlight and Deep Learning with Yacine decks
Co-authored-by: Cursor <cursoragent@cursor.com>e75c23e