{ "version": "run2b-v1", "config": "C0L", "stage": "C0", "max_len": 320, "base": "microsoft/deberta-v3-large", "input": ":\n C0a C0 + A3 citation-deleted copies of F's citing sentences as candidates (tagged non-citing, nearest OTHER\n citation as context, gold label kept) + paraphrase swap p=0.3. Otherwise identical to C0 (deberta-v3-base).\n C0L C0 with microsoft/deberta-v3-large: lr 1e-5, micro-batch 4 x 4 accumulation (= 16), gradient\n checkpointing (run 2a ran out of memory at micro-batch 8 without it on a 15 GB T4).\nRobustness (run 2a lost 50 jobs to one OOM): a failed job is logged (\"FAILED\"), marked in failed/, its GPU memory\nreleased, and the worker carries on with other jobs; re-running launch retries failed jobs.\nFinal model per config (all 15 papers, seed 1) -> models// and test_SEALED/.json (do not open).\nEnv: R1_SEEDS R1_CONFIGS R1_EPOCHS R1_BASE R1_OUT; R1_SMOKE=1 = quick check (2 folds, 1 seed, 1 epoch, 40 sentences).", "decision": "per-article z-score of the logit >= threshold (threshold from phase8.py)" }