FINER-SQL Arctic-Text2SQL-R1 reward contrast — 0.5B, BIRD train only
Final checkpoints (global_step_148, end of 1 epoch / 147 steps) from the matched-pair
Arctic-Text2SQL-R1 reward contrast documented in
campaigns/chained_05b_tkde/arctic_05b_bird/RESULTS_ARCTIC05B.md (finer-sql-tkde repo,
paper-and-architecture branch). Two arms, everything held fixed except which database the
reward executes on.
Base model: Qwen-2.5-Coder-0.5B-SQL-Writer (the FINER-SQL SFT writer checkpoint, NOT the
released RL-trained FINER-SQL 0.5B).
Config (Arctic's reward, campaign's fast 0.5B geometry): reward = Arctic's 1 / 0.1 / 0 set()
comparator (no format term) plus the statement gate from the 2026-09-15 collapse fix; 64 prompts
x 4 rollouts (256 sequences/step); beta 0.001 with the reference model on; ARCTIC_VARIANT=eager;
lr 8e-6; temperature 1.0; clip 0.2; entropy_coeff 0; completion length 2048; 1 epoch = 147 steps
(total_training_steps 148) over all 9,428 BIRD-train prompts. SQL executed only through the
db_execution API.
| folder | arm | reward database (tree) |
|---|---|---|
arctic05b_coredb/global_step_148 |
coredb | the frozen CoRE-DB BIRD tree (compact, enriched) |
arctic05b_origdb/global_step_148 |
origdb | this campaign's own copy of the original BIRD train DBs |
EX result (from the ledger, native rule-C extraction, BIRD dev, n=1534)
| arm | EX @ step 148 |
|---|---|
| origdb | 0.2% |
| coredb | 14.93% |
| headline gap | +14.73 pp |
Peak checkpoint: origdb peaked earlier (step 37, not end-of-epoch); coredb peaked at end-of-epoch
(step 148, this upload). Every EX row in the source ledger is reported twice (native format vs
Arctic's literal fence) — see RESULTS_ARCTIC05B.md for the full table and the reading protocol
(review/GATE5_PROTOCOL.md in the same directory). Document md5s of record are listed in
OPS_HANDOFF_ARCTIC05B.md.
This is not a reproduction of Arctic's own published EX — see the campaign README for the scope caveat (Snowflake never released their training code/data/geometry).