CoRE3B recovery checkpoints

Public resumable backups for one user-authorized3000-step run, not an accuracy-validated release. Initialization: released FINER-SQL-3B-BIRD. Fixed9687 sampled training examples (5100BIRD+4587Spider), all215 CoRE DBs, beta0/reward5/memory+atomic enabled. Fresh optimizer and memory; no matched ORI3B control.

Every500 steps (plus early recovery step20): model shards, optimizer, scheduler, RNG, trainer state, drained step-aligned memory. Restore checkpoint and sibling memory together. Full1534BIRD/1034Spider greedy evaluation at2000/3000. These are separate from the0.5B k30+VAV comparison.

Input limitations are recorded in RUN_SPEC.json:550 rows have one CoRE state;29 rows have all-NULL gold on selected CoRE states. No row is silently dropped and no missing DB falls back to ORI. Generic FROM parsing is repaired; Spider reward executes the SQL exactly as supplied, with no DISTINCT rewriting.

Downloads last month: -; Downloads are not tracked for this model. How to track

Video Preview

Reinforcement Learning

Model tree for thanhdath/FINER-SQL-3B-CoRE-R2-RL3000-runpod-20261023

Base model

griffith-bigdata/FINER-SQL-3B-BIRD

Finetuned

(1)

this model