thanhdath/mats-selection-think-review

MATS-SQL — Selection-3B THINK (reasoning duel judge) data review

Pairwise CHASE-SQL-style selection judge, THINK variant. Base = Llama-3.2-3B-Instruct, SFT-only (no ORPO). Completion = a ≤5-sentence reasoning chain (from gpt-oss-120b, hint-conditioned then hint-removed) followed by a final line Answer: A or Answer: B. Decoded at temperature 0.0.

Why THINK: the verdict-only judge (completion = a single A/B) beats majority-vote but is stuck at its conversion frontier — failure-close data augmentation was a wash. A 1-token judge has no channel to reason about why a small/empty result is correct. This dataset adds that channel.

Row schema

prompt = the duel (GRAST-30 schema + question + evidence + two candidate SQLs with 5-row execution previews). completion = reasoning + Answer: X. correct_letter = the ground-truth better candidate. src ∈ {bird, spider, synsql}, pairtype ∈ {cw = correct-vs-wrong, bw = both-wrong-closer-to-gold}.

Composition (train, by src × pairtype)

src	pairtype	n
bird	bw	466
bird	cw	5089
spider	bw	521
spider	cw	3420
synsql	bw	253
synsql	cw	3782

Quality gates (measured — review these)

gate	train	test
rows	13531	516
reasoning concludes the CORRECT letter (Answer: == correct_letter)	100.00%	100.00%
verdict parseable	100.00%	100.00%
A / B balance	6670 / 6861	257 / 259
hint-leak mentions	9	0
reasoning length (chars) p50 / p95	652 / 859	—

Train/test are DB-disjoint (held-out databases), so heldout letter-acc is not memorised.

Training config

• Base: meta-llama/Llama-3.2-3B-Instruct — full fine-tune (no LoRA/PEFT)
• Loss: completion-only (reasoning + verdict; prompt masked to −100)
• Epochs: 2, checkpoint saved every epoch (-ep0, -ep1); pick best by heldout letter-acc
• Effective batch: 64 (micro-batch 4 × grad-accum 16)
• LR: 1e-5, cosine schedule, 3% warmup, AdamW (weight_decay 0), grad-clip 1.0
• max_len: 3072, bf16, gradient checkpointing on
• Inference: temperature 0.0; round-robin/winner-stays tournament over the top-3 majority-vote clusters, τ=0.80 confidence gate over MV

Examples

Example 1 — src=bird, pairtype=cw, correct=A

Reasoning + verdict (completion):

Candidate A joins the production table (which links each car ID to a country) with country and counts the matching production rows, thus counting every car whose origin country is ‘Europe’. Candidate B only counts rows in the country lookup table where the name is ‘Europe’, which yields the number of country records, not the number of cars. The result of A (102) reflects a car‑level count, whereas B’s result (1) reflects a single country entry. Therefore, A correctly answers “how many cars … originated from Europe”.
Answer: A

Example 2 — src=synsql, pairtype=bw, correct=B

Reasoning + verdict (completion):

Candidate A joins the three tables without aggregation, so each cash_prizes row produces a separate result row, leading to duplicate event rows and mixing of highest_score values (e.g., 850.0 appears with the same event). It also reports only the individual prize amount rather than the total prize money awarded for the event. Candidate B groups by the event fields and sums the cash_prizes.amount, producing one row per event with the correct total prize money (330.0) and the appropriate highest_score. The ordering in both queries follows the required date descending then prize money descending, but only B fulfills the request for total prize money per completed event. Therefore, Candidate B better answers the question.
Answer: B

Example 3 — src=spider, pairtype=cw, correct=B

Reasoning + verdict (completion):

Both candidates join Rating to Movie on the movie identifier and group by the movie to compute an average of the star ratings. The only syntactic difference is the ORDER BY clause: Candidate A orders the averages in descending order (`ORDER BY avg(stars) DESC`) and then limits to one row, which would return the highest‑rated movie, whereas Candidate B orders the averages in ascending order (`ORDER BY avg(rating.stars)`) and limits to one row, which correctly selects the movie with the lowest average rating. The execution results illustrate this difference: Candidate A’s result shows a high average (4.5) for “Snow White,” while Candidate B’s result shows a low average (2.5) for “The Sound of Music.” Since the question asks for the titles and average ratings of movies with the lowest average rating, the ascending order used by Candidate B aligns with the requirement, whereas Candidate A does not. Therefore, Candidate B better answers the question.
Answer: B