harshbhx/ratchet-nemotron-nano-4b-text2sql

Text Generation • Updated 9 days ago • 118 • 2

Note Start here. The headline result: 39.0% and 33.0% on two unseen SQL exams (untrained 12.8% and 12.5%).

harshbhx/ratchet-nemotron-nano-4b-text2sql-nebius

Text Generation • Updated 9 days ago • 46

Note Round-2 adapter retrained on Nebius AI Cloud (spot L40S); reproduces 40.7% / 34.5%

Note SQL training data: 4,322 synthetic rows labelled by Nemotron 3 Ultra.

Note SQL exams: 172 + 200 questions the adapter never saw. Evaluation only; every answer saved.

harshbhx/ratchet-nemotron-nano-4b-toolcall

Text Generation • Updated 9 days ago • 107

Note Second result: 90.7% on unseen tools (untrained 62.7%); BFCL 80.3% (77.3%). Strict scoring.

Note Tool-calling training data: 2,940 synthetic rows labelled by Nemotron 3 Ultra.

Note Hard exam: 150 rows, 20 tools never seen in training. Evaluation only.

Note BFCL exam: 300 rows from the Berkeley Function Calling Leaderboard (Apache-2.0).

Note Try it: browse every exam row, untrained, adapter and teacher answers side by side.

Note Scout: three recorded live research runs with real web searches.

Note Try it: replay the recorded Scout runs at their real pace.