ASDCZX12DS/MyAwesomeModel-best

MyAwesomeModel (best checkpoint)

Selected checkpoint: step_1000 — the checkpoint with the highest weighted evaluation score across all 15 benchmarks (overall score = 0.710).

Evaluation Results

Scores below are reported to 3 decimal places.

Per-benchmark scores for the selected checkpoint (step_1000)

Benchmark	Score
math_reasoning	0.550
code_generation	0.650
text_classification	0.828
sentiment_analysis	0.792
question_answering	0.607
logical_reasoning	0.819
common_sense	0.736
reading_comprehension	0.700
dialogue_generation	0.644
summarization	0.767
translation	0.804
knowledge_retrieval	0.676
creative_writing	0.610
instruction_following	0.758
safety_evaluation	0.739
Overall (weighted)	0.710

All checkpoints — full benchmark matrix

Checkpoint	math_reasoning	code_generation	text_classification	sentiment_analysis	question_answering	logical_reasoning	common_sense	reading_comprehension	dialogue_generation	summarization	translation	knowledge_retrieval	creative_writing	instruction_following	safety_evaluation	Overall
step_100	0.345	0.350	0.517	0.617	0.475	0.319	0.530	0.475	0.438	0.517	0.640	0.529	0.328	0.550	0.628	0.480
step_200	0.383	0.421	0.603	0.675	0.510	0.375	0.583	0.529	0.493	0.600	0.700	0.570	0.388	0.610	0.650	0.535
step_300	0.415	0.475	0.667	0.710	0.533	0.445	0.621	0.569	0.530	0.650	0.733	0.596	0.436	0.650	0.668	0.576
step_400	0.443	0.517	0.714	0.733	0.550	0.525	0.650	0.600	0.557	0.683	0.755	0.615	0.475	0.679	0.683	0.608
step_500	0.467	0.550	0.750	0.750	0.564	0.605	0.672	0.625	0.579	0.707	0.769	0.631	0.507	0.700	0.696	0.635
step_600	0.487	0.577	0.776	0.762	0.575	0.675	0.690	0.645	0.596	0.725	0.780	0.643	0.534	0.717	0.707	0.656
step_700	0.506	0.600	0.795	0.772	0.584	0.731	0.705	0.663	0.611	0.739	0.788	0.653	0.557	0.730	0.717	0.674
step_800	0.522	0.619	0.809	0.780	0.593	0.773	0.717	0.677	0.624	0.750	0.795	0.662	0.577	0.741	0.725	0.689
step_900	0.537	0.636	0.820	0.786	0.600	0.801	0.727	0.689	0.634	0.759	0.800	0.670	0.595	0.750	0.732	0.700
step_1000	0.550	0.650	0.828	0.792	0.607	0.819	0.736	0.700	0.644	0.767	0.804	0.676	0.610	0.758	0.739	0.710

How the scores are computed

Each benchmark score is a deterministic function of the training step number (evaluated via the workspace evaluation/ suite), and the overall score is their weighted average (weights defined in evaluation/eval.py). All 15 benchmark functions are monotonically increasing with training step, so the latest checkpoint step_1000 achieves the highest overall score.