Note Canonical TRAINING suite: 2238 Harbor data-analysis tasks over real Kaggle datasets. Single correctness reward.

Note Held-out EVAL suite: 366 verified Harbor tasks. For benchmarking, not training.

Note Source-of-truth row-level split manifest (parquet, eval/train, ~30k rows). The Harbor task suites are built from this.

Note Shared Kaggle DATA BUCKET every task pulls from at container start (keyed by BUCKET_PREFIX).

Note Fast-iteration SUBSET: 100 easy (L1) numeric-reward tasks sampled from the train suite.

Note Train suite (2238) with a 3-REWARD verifier: correctness + submission + tool_efficiency (reward.json).

Note Train suite (2238) ORDERED easy->hard by empirical pass@4 difficulty; registry carries rank/difficulty/solve_frac. Curriculum-ready + multireward.

AdithyaSK/data-agent-2b-normal-final

Text Generation • 3B • Updated 1 day ago • 38

Note 2B · normal · FINAL (step 1119) — pass@4 0.603 vs base 0.284 (~2.1×), pass@1 0.41. GRPO on Harbor data-agent.

AdithyaSK/data-agent-2b-normal-best

Text Generation • 3B • Updated 1 day ago • 27

Note 2B · normal · BEST-eval (step 800) — peak pass@4 0.624.

AdithyaSK/data-agent-2b-curriculum-final

Text Generation • 3B • Updated 1 day ago • 26

Note 2B · curriculum (easy→hard) · FINAL (step 1119) — pass@4 0.603. Same endpoint as normal.

AdithyaSK/data-agent-2b-curriculum-best

Text Generation • 3B • Updated 1 day ago • 28

Note 2B · curriculum · BEST-eval (step 900) — peak pass@4 0.606.

AdithyaSK/data-agent-4b-normal-final

Text Generation • 5B • Updated 1 day ago • 23

Note 4B · normal · FINAL (step 1119) — pass@4 0.791 vs base 0.740, pass@1 0.64. Strong base, modest lift.

AdithyaSK/data-agent-4b-normal-best

Text Generation • 5B • Updated 1 day ago • 24

Note 4B · normal · BEST-eval (step 200) — peak pass@4 0.802; learned early, plateaued.