2 days ago
-
Refresh sft and all with the final nightly top-ups (58,424 verified traces, 131,227 generations)
3 days ago
-
Add Beam's own eval answers at every reasoning effort and the effort-dial study
9 days ago
-
v2: Beam reasoning-effort levels (low/medium/high), Accurate variant; v1 is tagged
-
Re-evaluate at a 64k-token limit: resume answers cut off at 32k (corrects AIME comparisons)
11 days ago
-
Switch all files to Parquet so every config (incl. eval) works in the viewer; same rows
-
Add eval outputs: every graded answer behind the Dory-Beam-4B results (GPQA text omitted)
-
Beam reasoning traces: 53,306 verified SFT rows + 124,945 raw generations
-
initial commit
