Datasetpublic

tts-sft/round5-abcd-cells

Round-5 self-verifier D-cut — four new-verifier mechanisms (2026-08-18) Four opt-in verifier mechanisms on top of the v7 stack (official-sample anchoring + validity probes + wb-cands 8 + wb-certify), one arm each, on the 295-problem mechanism-screening slice (u00+u01 of the 589 pool; same problems, budgets, loop-0 population and GENSEED as the observed sv cells — rows are directly comparable to the sv7/sv7d/pw7/sel7/sum7 screen table). arm flag mechanism bru7… See the full description on the dataset page: https://huggingface.co/datasets/tts-sft/round5-abcd-cells.

sourceHugging Faceotherupdated 1mo agoView on Hugging Face

0likes80downloads

Dataset Card

Round-5 self-verifier D-cut — four new-verifier mechanisms (2026-08-18)

Four opt-in verifier mechanisms on top of the v7 stack (official-sample anchoring + validity probes + wb-cands 8 + wb-certify), one arm each, on the 295-problem mechanism-screening slice (u00+u01 of the 589 pool; same problems, budgets, loop-0 population and GENSEED as the observed sv cells — rows are directly comparable to the sv7/sv7d/pw7/sel7/sum7 screen table).

arm	flag	mechanism
`bru7`	`--brute-ref`	A: sample-gated BRUTE-FORCE reference joins wb certification on small inputs (brute + 1 efficient ref = certified; the classic stress-test triangle's missing leg)
`rdx7`	`--ref-diverse 3`	B: three sample-gated re-reading references; where sample-consistent re-reads disagree with held labels, the DISPUTE (not a manufactured label) is surfaced via the summary channel (loops 1-4)
`pv7`	`--out-validators`	C: any-valid-answer problems get a certified OUTPUT-PROPERTY validator (gate: accepts official sample outputs, rejects corrupted ones); labels stripped, tests become property-checked instead of skipped
`tr7`	`--trace-repair`	D: problems with >=50% of the population blocked for >=2 straight loops switch the next repair round to hand-trace mode via the summary channel

Contents: units_cells/ (8 units + manifest) · run_shard_r5_sva.sh (same contract as run_shard_r5_sv.sh: env UNIT + RUN_DIR) · minipatch/ (sv minipatch + the D-cut patches to selffb/selffb_v2_prepare.py and scripts/lcb_public_probe_harness_spj.py; all mechanisms are opt-in flags — behavior without the flags is bit-identical, verified) · validate_abcd.py (17 CPU tests, stubbed-LLM integration through the real pipeline; run TMPDIR=/var/tmp python3 validate_abcd.py after cloning, needs no GPU).

Return the whole RUN_DIR per unit (especially selffb/, ck_nonsat/, fb_audit.jsonl, metrics_loops.json, *.summary.jsonl); grading happens on our side. New funnel fields to expect in the prepare audits: n_brute_valid, n_wb_cert_brute, n_diverse_valid, n_disputes, oval_certified, trace_blocked_frac, trace_streak.