{ "version": "1.0.0", "notebook_version": "V3.7", "mode": "scratch", "parent_version": null, "lineage": [ "jhu-clsp/ettin-encoder-150m" ], "config": { "MODE": "scratch", "FINETUNE_FROM": "latest", "MODEL_NAME": "LightDec_V2_Long", "BACKBONE": "jhu-clsp/ettin-encoder-150m", "BACKBONE_INIT": "pretrained", "PRESET": "long", "CUSTOM_DATA_JSONL": "", "CUSTOM_WEIGHT": 3.0, "INCLUDE_BUILDERS": [], "EXCLUDE_BUILDERS": [ "mind2web" ], "TEV1_DIR": "./tev1", "TEV1_AUTOBUILD": true, "TEV1_POLICY_CAP": 13500, "TEV1_ROUTING_CAP": 6000, "TEV1_RESEARCH_CAP": 1000, "TRIAGE_EMAILS": 200, "TEV1_RETRIES": 3, "TEV1_STEP_TIMEOUT_MIN": 180, "TEV1_STALL_MIN": 15, "RETRY_FAILED_BUILDERS": true, "LONG_TASK_TRAIN": 1500, "MAX_LEN": 2048, "LONG_MAX_LEN": 2048, "LONG_OPTS_THRESHOLD": 24, "HEAD_MAX_LEN": 192, "MAX_TOK_PER_OPT": 24, "MAX_OPTS_SINGLE_PASS": 96, "INTERACT_LAYERS": 2, "INTERACT_HEADS": 8, "BATCH_SIZE": 128, "TOKENS_PER_BATCH": 65536, "GRAD_ACCUM": 1, "LR_ENCODER": 8e-05, "LR_HEAD": 0.0006, "LLRD": 0.9, "WEIGHT_DECAY": 0.01, "WARMUP_FRAC": 0.06, "EPOCHS": 0, "EMA_DECAY": 0.999, "GRAD_CLIP": 1.0, "FINETUNE_LR_SCALE": 0.5, "SPHERICAL_W": 0.5, "BRIER_W": 0.0, "RPS_W": 1.0, "TASK_SAMPLING_ALPHA": 0.5, "NOTA_PROB": 0.08, "RLCD_EPOCHS": 0.0, "RLCD_SAMPLES": 4, "RLCD_SIGMA": 0.3, "RLCD_CE_ANCHOR": 0.5, "DEVICE": "cuda", "SEED": 42, "NUM_WORKERS": 8, "USE_COMPILE": true, "ATTN_IMPL": "auto", "OUTPUT_ROOT": "./falcondec_runs", "EXPORT_INT8": true, "SIZE_TARGET_MB": 250, "DEFER_THRESHOLD": 0.7, "EVAL_BASELINE_PROOF_V2": true, "BASELINE_MAX_PER_TASK": 120, "RUN_CPU_BENCH": true, "PUSH_TO_HUB": false, "HUB_REPO": "" }, "epochs": 10, "best_epoch": 4, "rlcd_kept": false, "history": [ { "epoch": 1, "train_loss": 0.7058714814313827, "train_acc": 0.7342094450703411, "val_macro": 0.8188568656508199, "val_micro": 0.820702479338843, "val_nll": 0.4076740937270036 }, { "epoch": 2, "train_loss": 0.431469622218984, "train_acc": 0.8473141073247583, "val_macro": 0.8450872188411825, "val_micro": 0.8444628099173553, "val_nll": 0.3623910443100994 }, { "epoch": 3, "train_loss": 0.34276796075090776, "train_acc": 0.8791940521680108, "val_macro": 0.8501904675940947, "val_micro": 0.8499173553719008, "val_nll": 0.39227130460368825 }, { "epoch": 4, "train_loss": 0.2773282952523419, "train_acc": 0.9039863196679513, "val_macro": 0.8519172452047367, "val_micro": 0.8519834710743802, "val_nll": 0.4529550437184426 }, { "epoch": 5, "train_loss": 0.22030453415052323, "train_acc": 0.925851485588573, "val_macro": 0.8488092001804082, "val_micro": 0.85, "val_nll": 0.5369876145373608 }, { "epoch": 6, "train_loss": 0.16953414903032543, "train_acc": 0.9449156076226879, "val_macro": 0.8474279008050457, "val_micro": 0.8488842975206612, "val_nll": 0.6705887193589142 }, { "epoch": 7, "train_loss": 0.126976966611751, "train_acc": 0.9602833527848939, "val_macro": 0.8453322694829587, "val_micro": 0.8476446280991735, "val_nll": 0.8430423012207805 }, { "epoch": 8, "train_loss": 0.09470793687766035, "train_acc": 0.9714013975720073, "val_macro": 0.8451769009269958, "val_micro": 0.8467355371900827, "val_nll": 1.0336680755654943 }, { "epoch": 9, "train_loss": 0.0710737590306488, "train_acc": 0.9794225926173407, "val_macro": 0.8435159403610613, "val_micro": 0.8458677685950413, "val_nll": 1.210751206553627 }, { "epoch": 10, "train_loss": 0.05665995772408678, "train_acc": 0.9843832093140154, "val_macro": 0.844272704258198, "val_micro": 0.8474380165289256, "val_nll": 1.41872131482475 } ], "train_minutes": 291.67019440333047, "temperatures": [ [ 2.37591814994812, 2.0426506996154785, 1.6455482244491577, 1.8417285680770874 ], [ 2.4885058403015137, 2.4885058403015137, 2.4885058403015137, 2.4885058403015137 ], [ 2.2925639152526855, 2.2925639152526855, 2.2925639152526855, 2.2925639152526855 ] ], "data_counts": { "train": 1023814, "validation": 24200, "test": 31990 }, "skipped_builders": [], "test_overall": { "n": 31990.0, "micro_acc": 0.784057517974367, "macro_acc": 0.7887006006875437, "nll": 0.5630830966790886, "brier": 0.2954194093899411, "ece": 0.02856345933167103, "aurc": 0.06778106632931055, "score_mae": 0.5060575584353655, "coverage@0.7": 0.6797436698968428, "acc_on_covered@0.7": 0.9059094044607956 }, "test_per_domain": { "agentic": 0.821117397177404, "classification": 0.6115, "code": 0.9108344674425007, "guardrails": 0.7984073149310548, "intents": 0.8505, "long_context": 1.0, "policy": 0.8061818181818182, "reasoning": 0.7180877154615182, "support": 0.9815181518151815, "tev1_benchmark": 0.6541666666666667, "workflows": 0.773 }, "test_per_task": [ { "task": "ag_news/topic", "n": 500, "acc": 0.9, "chance": 0.25, "nll": 0.2674782864307199, "brier": 0.14296993909824685, "heldout": false, "domain": "classification", "ece": 0.05165432274341582, "proof_v2 (card)": NaN, "\u0394 vs proof_v2": NaN }, { "task": "agentharm/refuse", "n": 416, "acc": 0.6995192307692307, "chance": 0.5, "nll": 0.6008703433729422, "brier": 0.4076762512706606, "heldout": true, "domain": "guardrails", "ece": 0.08916855761064935, "proof_v2 (card)": NaN, "\u0394 vs proof_v2": NaN }, { "task": "agenttrek/finish_now", "n": 151, "acc": 0.7880794701986755, "chance": 0.5, "nll": 0.445978463758558, "brier": 0.30066953632150406, "heldout": false, "domain": "agentic", "ece": 0.07512128787324918, "proof_v2 (card)": NaN, "\u0394 vs proof_v2": NaN }, { "task": "agenttrek/next_action_type", "n": 487, "acc": 0.864476386036961, "chance": 0.19054952576513148, "nll": 0.3929848723152591, "brier": 0.2184262054212701, "heldout": false, "domain": "agentic", "ece": 0.09202881025827395, "proof_v2 (card)": NaN, "\u0394 vs proof_v2": NaN }, { "task": "anli/nli", "n": 498, "acc": 0.4859437751004016, "chance": 0.33333333333333326, "nll": 1.069221544292677, "brier": 0.6544603492386665, "heldout": false, "domain": "reasoning", "ece": 0.14589204983299517, "proof_v2 (card)": NaN, "\u0394 vs proof_v2": NaN }, { "task": "aqua_rat/math", "n": 247, "acc": 0.340080971659919, "chance": 0.2, "nll": 1.54525283087603, "brier": 0.7739737508966391, "heldout": false, "domain": "reasoning", "ece": 0.080350092668765, "proof_v2 (card)": NaN, "\u0394 vs proof_v2": NaN }, { "task": "arc_challenge/mcq", "n": 500, "acc": 0.49, "chance": 0.25006666666666666, "nll": 1.3021090898208785, "brier": 0.6752390805252713, "heldout": true, "domain": "reasoning", "ece": 0.1680490539073944, "proof_v2 (card)": 0.308, "\u0394 vs proof_v2": 0.182 }, { "task": "arc_easy/mcq", "n": 500, "acc": 0.612, "chance": 0.24980000000000002, "nll": 0.9830388274090365, "brier": 0.5264415342806388, "heldout": true, "domain": "reasoning", "ece": 0.1215095258355141, "proof_v2 (card)": 0.425, "\u0394 vs proof_v2": 0.187 }, { "task": "banking77/intent", "n": 500, "acc": 0.922, "chance": 0.3214666666666666, "nll": 0.23128961177584006, "brier": 0.11756106243365176, "heldout": true, "domain": "intents", "ece": 0.037492330908775316, "proof_v2 (card)": 0.883, "\u0394 vs proof_v2": 0.039000000000000035 }, { "task": "banking77/intent_77", "n": 300, "acc": 0.58, "chance": 0.012987012987012986, "nll": 2.159930980304489, "brier": 0.6431857202493357, "heldout": true, "domain": "intents", "ece": 0.24079640249411272, "proof_v2 (card)": NaN, "\u0394 vs proof_v2": NaN }, { "task": "bigclonebench/clone", "n": 500, "acc": 0.962, "chance": 0.5, "nll": 0.1312969167594565, "brier": 0.06444151702850223, "heldout": false, "domain": "code", "ece": 0.03135248208045959, "proof_v2 (card)": 0.383, "\u0394 vs proof_v2": 0.579 }, { "task": "bitext/category", "n": 500, "acc": 1.0, "chance": 0.16128124098124094, "nll": 0.0013673773490934594, "brier": 0.0003828941821233322, "heldout": false, "domain": "support", "ece": 0.001247308015823329, "proof_v2 (card)": NaN, "\u0394 vs proof_v2": NaN }, { "task": "bitext/route", "n": 500, "acc": 1.0, "chance": 0.19016313131313134, "nll": 0.00031508885690954, "brier": 4.903548307607296e-06, "heldout": false, "domain": "support", "ece": 0.0003137041330337764, "proof_v2 (card)": 0.958, "\u0394 vs proof_v2": 0.04200000000000004 }, { "task": "boolq/yes_no", "n": 500, "acc": 0.822, "chance": 0.5, "nll": 0.4842173567947466, "brier": 0.2840138305161832, "heldout": false, "domain": "reasoning", "ece": 0.0795972980260849, "proof_v2 (card)": 0.717, "\u0394 vs proof_v2": 0.10499999999999998 }, { "task": "civil_comments/toxic", "n": 500, "acc": 0.926, "chance": 0.5, "nll": 0.2234244500696659, "brier": 0.1209681824839501, "heldout": false, "domain": "guardrails", "ece": 0.05123518884181974, "proof_v2 (card)": NaN, "\u0394 vs proof_v2": NaN }, { "task": "clinc150/intent", "n": 500, "acc": 0.97, "chance": 0.12404474984829318, "nll": 0.09906445816022005, "brier": 0.04996487012600131, "heldout": false, "domain": "intents", "ece": 0.014944501101970714, "proof_v2 (card)": 0.85, "\u0394 vs proof_v2": 0.12 }, { "task": "codexglue/code_to_doc", "n": 504, "acc": 0.9900793650793651, "chance": 0.26233465608465617, "nll": 0.032474144255576765, "brier": 0.016158895201785088, "heldout": false, "domain": "code", "ece": 0.009654374349685009, "proof_v2 (card)": 0.969, "\u0394 vs proof_v2": 0.021079365079365142 }, { "task": "codexglue/doc_to_code", "n": 504, "acc": 0.9761904761904762, "chance": 0.2786044973544973, "nll": 0.05970113779317181, "brier": 0.03072688989396561, "heldout": false, "domain": "code", "ece": 0.012547755761752008, "proof_v2 (card)": 0.961, "\u0394 vs proof_v2": 0.015190476190476199 }, { "task": "codexglue/func_name", "n": 467, "acc": 0.9550321199143469, "chance": 0.25499643112062814, "nll": 0.14961532151979784, "brier": 0.07053876912275295, "heldout": false, "domain": "code", "ece": 0.03948740715388316, "proof_v2 (card)": 0.901, "\u0394 vs proof_v2": 0.054032119914346866 }, { "task": "codexglue/lang_id", "n": 504, "acc": 1.0, "chance": 0.2354166666666667, "nll": 0.0014352430686462487, "brier": 1.2864369183890018e-05, "heldout": false, "domain": "code", "ece": 0.0014309174721203188, "proof_v2 (card)": 0.997, "\u0394 vs proof_v2": 0.0030000000000000027 }, { "task": "commonsense_qa/mcq", "n": 493, "acc": 0.6430020283975659, "chance": 0.19999999999999998, "nll": 0.9523430099929484, "brier": 0.49435087368161623, "heldout": false, "domain": "reasoning", "ece": 0.14308575696442233, "proof_v2 (card)": 0.442, "\u0394 vs proof_v2": 0.20100202839756592 }, { "task": "counsel/critique_quality", "n": 201, "acc": 0.6268656716417911, "chance": 0.3333333333333333, "nll": 1.5466346353110312, "brier": 0.6427543071457474, "heldout": false, "domain": "agentic", "ece": 0.27062572442477023, "proof_v2 (card)": NaN, "\u0394 vs proof_v2": NaN }, { "task": "counsel/step_has_error", "n": 201, "acc": 0.8308457711442786, "chance": 0.5, "nll": 0.8236887436017318, "brier": 0.3201374357369142, "heldout": false, "domain": "agentic", "ece": 0.14827988011326954, "proof_v2 (card)": NaN, "\u0394 vs proof_v2": NaN }, { "task": "devign/vulnerability", "n": 500, "acc": 0.632, "chance": 0.5, "nll": 0.6074751503933221, "brier": 0.42706214521965163, "heldout": false, "domain": "code", "ece": 0.07018342161178594, "proof_v2 (card)": 0.542, "\u0394 vs proof_v2": 0.08999999999999997 }, { "task": "emotion/6way", "n": 500, "acc": 0.472, "chance": 0.16666666666666663, "nll": 1.561316128242761, "brier": 0.7356062387487318, "heldout": true, "domain": "classification", "ece": 0.21054179659485817, "proof_v2 (card)": NaN, "\u0394 vs proof_v2": NaN }, { "task": "gsm8k/math", "n": 500, "acc": 0.7, "chance": 0.25, "nll": 0.6923719964642078, "brier": 0.4014884319624292, "heldout": false, "domain": "reasoning", "ece": 0.08541783547401426, "proof_v2 (card)": 0.275, "\u0394 vs proof_v2": 0.42499999999999993 }, { "task": "hellaswag/continuation", "n": 500, "acc": 0.576, "chance": 0.25, "nll": 0.968790470642969, "brier": 0.5288133586177948, "heldout": false, "domain": "reasoning", "ece": 0.07543198531866073, "proof_v2 (card)": NaN, "\u0394 vs proof_v2": NaN }, { "task": "hotpotqa/comparison_yes_no", "n": 26, "acc": 0.9230769230769231, "chance": 0.5, "nll": 0.3400710394176153, "brier": 0.15484679762361436, "heldout": false, "domain": "agentic", "ece": 0.08232885369887719, "proof_v2 (card)": NaN, "\u0394 vs proof_v2": NaN }, { "task": "hotpotqa/retrieve", "n": 497, "acc": 0.8933601609657947, "chance": 0.16741720162243304, "nll": 0.3457089332123268, "brier": 0.16216258689991922, "heldout": false, "domain": "agentic", "ece": 0.04268557528854612, "proof_v2 (card)": NaN, "\u0394 vs proof_v2": NaN }, { "task": "humaneval/completion", "n": 119, "acc": 0.8487394957983193, "chance": 0.39355742296918783, "nll": 0.46361927195851294, "brier": 0.2730045795813753, "heldout": true, "domain": "code", "ece": 0.13989568708323633, "proof_v2 (card)": 0.575, "\u0394 vs proof_v2": 0.27373949579831935 }, { "task": "jailbreak/detect", "n": 262, "acc": 0.9732824427480916, "chance": 0.5, "nll": 0.08469655378148465, "brier": 0.044003931208438284, "heldout": false, "domain": "guardrails", "ece": 0.030764950368240628, "proof_v2 (card)": NaN, "\u0394 vs proof_v2": NaN }, { "task": "long/contract_clause", "n": 150, "acc": 1.0, "chance": 0.19999999999999996, "nll": 0.0002789537957808837, "brier": 4.274148862877054e-07, "heldout": false, "domain": "long_context", "ece": 0.00027882695198055973, "proof_v2 (card)": NaN, "\u0394 vs proof_v2": NaN }, { "task": "long/email_thread", "n": 150, "acc": 1.0, "chance": 0.25, "nll": 0.0003266528345905802, "brier": 2.717841913247229e-06, "heldout": false, "domain": "long_context", "ece": 0.0003259424368540209, "proof_v2 (card)": NaN, "\u0394 vs proof_v2": NaN }, { "task": "long/service_log", "n": 150, "acc": 1.0, "chance": 0.19999999999999996, "nll": 0.0010035930563268873, "brier": 2.0987077842670875e-06, "heldout": false, "domain": "long_context", "ece": 0.0010028688112894146, "proof_v2 (card)": NaN, "\u0394 vs proof_v2": NaN }, { "task": "massive_en/intent", "n": 500, "acc": 0.93, "chance": 0.12958727789085603, "nll": 0.20636169074449573, "brier": 0.10375277943487987, "heldout": false, "domain": "intents", "ece": 0.0284088468849659, "proof_v2 (card)": NaN, "\u0394 vs proof_v2": NaN }, { "task": "mbpp/bugspot", "n": 256, "acc": 0.85546875, "chance": 0.40559895833333326, "nll": 0.3667814989712497, "brier": 0.21843909652329055, "heldout": false, "domain": "code", "ece": 0.037241218378767364, "proof_v2 (card)": 0.475, "\u0394 vs proof_v2": 0.38046875 }, { "task": "mbpp/solution", "n": 500, "acc": 0.978, "chance": 0.25, "nll": 0.05863042902501184, "brier": 0.02981698831494921, "heldout": false, "domain": "code", "ece": 0.01960155099630357, "proof_v2 (card)": 0.992, "\u0394 vs proof_v2": -0.014000000000000012 }, { "task": "mmlu/mcq", "n": 500, "acc": 0.39, "chance": 0.25, "nll": 1.4543649319559335, "brier": 0.7702575193705765, "heldout": true, "domain": "reasoning", "ece": 0.15507307499647138, "proof_v2 (card)": NaN, "\u0394 vs proof_v2": NaN }, { "task": "mnli/claim", "n": 500, "acc": 0.86, "chance": 0.33333333333333326, "nll": 0.4006310784481466, "brier": 0.2131232579488494, "heldout": false, "domain": "reasoning", "ece": 0.081218329668045, "proof_v2 (card)": 0.492, "\u0394 vs proof_v2": 0.368 }, { "task": "openbookqa/mcq", "n": 500, "acc": 0.572, "chance": 0.25, "nll": 1.1920220527790952, "brier": 0.6155284183368397, "heldout": false, "domain": "reasoning", "ece": 0.21775564682483672, "proof_v2 (card)": 0.292, "\u0394 vs proof_v2": 0.27999999999999997 }, { "task": "policy/access_control_transfer", "n": 500, "acc": 1.0, "chance": 0.33333333333333326, "nll": 0.002006467206090747, "brier": 2.4728843418299015e-05, "heldout": false, "domain": "policy", "ece": 0.001999327063560541, "proof_v2 (card)": NaN, "\u0394 vs proof_v2": NaN }, { "task": "policy/count_threshold_transfer", "n": 500, "acc": 0.88, "chance": 0.10537052392052391, "nll": 0.319376940273738, "brier": 0.1809248790341789, "heldout": false, "domain": "policy", "ece": 0.04758520478010175, "proof_v2 (card)": NaN, "\u0394 vs proof_v2": NaN }, { "task": "policy/free_shipping_transfer", "n": 500, "acc": 0.938, "chance": 0.5, "nll": 0.1287872482436942, "brier": 0.08322479293591872, "heldout": false, "domain": "policy", "ece": 0.041308605194091734, "proof_v2 (card)": NaN, "\u0394 vs proof_v2": NaN }, { "task": "policy/invoice_overdue_transfer", "n": 500, "acc": 0.824, "chance": 0.5, "nll": 0.3965394942490384, "brier": 0.2504247577332596, "heldout": false, "domain": "policy", "ece": 0.01985593855381013, "proof_v2 (card)": NaN, "\u0394 vs proof_v2": NaN }, { "task": "policy/invoice_total_transfer", "n": 500, "acc": 0.474, "chance": 0.5, "nll": 0.710681935429573, "brier": 0.5164436840846652, "heldout": false, "domain": "policy", "ece": 0.08196303224563599, "proof_v2 (card)": NaN, "\u0394 vs proof_v2": NaN }, { "task": "policy/refund_approval_transfer", "n": 500, "acc": 0.962, "chance": 0.33333333333333326, "nll": 0.10006511305032836, "brier": 0.057969574647252664, "heldout": false, "domain": "policy", "ece": 0.018834196925163266, "proof_v2 (card)": NaN, "\u0394 vs proof_v2": NaN }, { "task": "policy/return_window_transfer", "n": 500, "acc": 1.0, "chance": 0.33333333333333326, "nll": 0.02710788446944207, "brier": 0.004185976644178241, "heldout": false, "domain": "policy", "ece": 0.02567652928829193, "proof_v2 (card)": NaN, "\u0394 vs proof_v2": NaN }, { "task": "policy/sla_urgency_transfer", "n": 500, "acc": 0.606, "chance": 0.25, "nll": 0.8170870101451874, "brier": 0.5123369210022903, "heldout": false, "domain": "policy", "ece": 0.13068743020296097, "proof_v2 (card)": NaN, "\u0394 vs proof_v2": NaN }, { "task": "policy/table_compare_transfer", "n": 500, "acc": 0.928, "chance": 0.5, "nll": 0.2105378701629961, "brier": 0.11898007306418568, "heldout": false, "domain": "policy", "ece": 0.04929606854915622, "proof_v2 (card)": NaN, "\u0394 vs proof_v2": NaN }, { "task": "policy/table_count_transfer", "n": 500, "acc": 0.324, "chance": 0.12114228549228548, "nll": 3.24442003638536, "brier": 1.1488026091975905, "heldout": false, "domain": "policy", "ece": 0.5407210965156555, "proof_v2 (card)": NaN, "\u0394 vs proof_v2": NaN }, { "task": "policy/table_extreme_transfer", "n": 500, "acc": 0.932, "chance": 0.12210815295815294, "nll": 0.268062693382595, "brier": 0.11322196827103552, "heldout": false, "domain": "policy", "ece": 0.05259446609020234, "proof_v2 (card)": NaN, "\u0394 vs proof_v2": NaN }, { "task": "prompt_injections/detect", "n": 116, "acc": 0.5948275862068966, "chance": 0.5, "nll": 1.1838965486195179, "brier": 0.6826645118148511, "heldout": true, "domain": "guardrails", "ece": 0.33272201953263114, "proof_v2 (card)": NaN, "\u0394 vs proof_v2": NaN }, { "task": "qasc/mcq", "n": 500, "acc": 0.986, "chance": 0.125, "nll": 0.028506858559036514, "brier": 0.017274368003708154, "heldout": false, "domain": "reasoning", "ece": 0.007774531126022318, "proof_v2 (card)": NaN, "\u0394 vs proof_v2": NaN }, { "task": "sciq/mcq", "n": 498, "acc": 0.9538152610441767, "chance": 0.25, "nll": 0.1506263591345834, "brier": 0.06985463059035983, "heldout": false, "domain": "reasoning", "ece": 0.020590302994452258, "proof_v2 (card)": 0.692, "\u0394 vs proof_v2": 0.26181526104417674 }, { "task": "scitail/support", "n": 500, "acc": 0.962, "chance": 0.5, "nll": 0.11612921302905306, "brier": 0.05990850379879065, "heldout": false, "domain": "reasoning", "ece": 0.03015338957309728, "proof_v2 (card)": NaN, "\u0394 vs proof_v2": NaN }, { "task": "snli/contradicts", "n": 500, "acc": 0.99, "chance": 0.33333333333333326, "nll": 0.042843449617153966, "brier": 0.017027529190616314, "heldout": false, "domain": "reasoning", "ece": 0.018511961221695038, "proof_v2 (card)": 0.892, "\u0394 vs proof_v2": 0.09799999999999998 }, { "task": "snli/must_be_true", "n": 500, "acc": 0.986, "chance": 0.33333333333333326, "nll": 0.06753946821717545, "brier": 0.02560326154020915, "heldout": false, "domain": "reasoning", "ece": 0.043202445626258815, "proof_v2 (card)": 0.908, "\u0394 vs proof_v2": 0.07799999999999996 }, { "task": "snli/nli", "n": 988, "acc": 0.8947368421052632, "chance": 0.33333333333333326, "nll": 0.3554420507991845, "brier": 0.17963376213248355, "heldout": false, "domain": "reasoning", "ece": 0.10791712172842222, "proof_v2 (card)": NaN, "\u0394 vs proof_v2": NaN }, { "task": "sst5/score", "n": 500, "acc": 0.406, "chance": 0.2, "nll": 1.3002369542717933, "brier": 0.6888741553408958, "heldout": true, "domain": "classification", "ece": 0.068851743131876, "proof_v2 (card)": NaN, "\u0394 vs proof_v2": NaN }, { "task": "tev1_test/ag_news", "n": 150, "acc": 0.92, "chance": 0.25, "nll": 0.2815313008365532, "brier": 0.13456139206846243, "heldout": true, "domain": "tev1_benchmark", "ece": 0.09770219445228576, "proof_v2 (card)": NaN, "\u0394 vs proof_v2": NaN }, { "task": "tev1_test/banking77", "n": 200, "acc": 0.72, "chance": 0.17791666666666664, "nll": 0.777236735031438, "brier": 0.3902663054271286, "heldout": true, "domain": "tev1_benchmark", "ece": 0.13721376925706863, "proof_v2 (card)": NaN, "\u0394 vs proof_v2": NaN }, { "task": "tev1_test/boolq", "n": 200, "acc": 0.835, "chance": 0.5, "nll": 0.40902234488283284, "brier": 0.24753055362669119, "heldout": true, "domain": "tev1_benchmark", "ece": 0.04955918818712238, "proof_v2 (card)": NaN, "\u0394 vs proof_v2": NaN }, { "task": "tev1_test/mnli", "n": 300, "acc": 0.7233333333333334, "chance": 0.3333333333333333, "nll": 0.6736717029288412, "brier": 0.39782655881245704, "heldout": true, "domain": "tev1_benchmark", "ece": 0.07533363938331603, "proof_v2 (card)": NaN, "\u0394 vs proof_v2": NaN }, { "task": "tev1_test/policy", "n": 1200, "acc": 0.5291666666666667, "chance": 0.3333333333333333, "nll": 0.8761954319352905, "brier": 0.5462664224307302, "heldout": true, "domain": "tev1_benchmark", "ece": 0.05868758827447889, "proof_v2 (card)": NaN, "\u0394 vs proof_v2": NaN }, { "task": "tev1_test/routing", "n": 600, "acc": 0.4583333333333333, "chance": 0.2, "nll": 1.120521725914441, "brier": 0.5789269511485964, "heldout": true, "domain": "tev1_benchmark", "ece": 0.037736288358767814, "proof_v2 (card)": NaN, "\u0394 vs proof_v2": NaN }, { "task": "tev1_test/sst5", "n": 150, "acc": 0.3933333333333333, "chance": 0.19999999999999996, "nll": 1.294863009850184, "brier": 0.6810328679494898, "heldout": true, "domain": "tev1_benchmark", "ece": 0.1081918982664744, "proof_v2 (card)": NaN, "\u0394 vs proof_v2": NaN }, { "task": "triage/support_email", "n": 505, "acc": 0.9445544554455445, "chance": 0.32333333333333325, "nll": 0.15705669974757994, "brier": 0.07848840114983771, "heldout": false, "domain": "support", "ece": 0.045869263741049465, "proof_v2 (card)": NaN, "\u0394 vs proof_v2": NaN }, { "task": "typed_decisions/agent_trace_observability", "n": 500, "acc": 0.734, "chance": 0.3, "nll": 0.8089503426551818, "brier": 0.45125534421065666, "heldout": false, "domain": "workflows", "ece": 0.22166652178764346, "proof_v2 (card)": NaN, "\u0394 vs proof_v2": NaN }, { "task": "typed_decisions/customer_service", "n": 500, "acc": 0.764, "chance": 0.28, "nll": 0.7428050636351109, "brier": 0.40177602517263195, "heldout": false, "domain": "workflows", "ece": 0.20972245779633528, "proof_v2 (card)": NaN, "\u0394 vs proof_v2": NaN }, { "task": "typed_decisions/invoice_processing", "n": 500, "acc": 0.826, "chance": 0.35, "nll": 0.5732684296518564, "brier": 0.2997603036136731, "heldout": false, "domain": "workflows", "ece": 0.19463082921504973, "proof_v2 (card)": NaN, "\u0394 vs proof_v2": NaN }, { "task": "typed_decisions/security_incidents", "n": 500, "acc": 0.768, "chance": 0.34, "nll": 0.7502468670606613, "brier": 0.4238190137148829, "heldout": false, "domain": "workflows", "ece": 0.23508695220947262, "proof_v2 (card)": NaN, "\u0394 vs proof_v2": NaN }, { "task": "winogrande/blank", "n": 500, "acc": 0.662, "chance": 0.5, "nll": 0.6651458515003323, "brier": 0.45323706188140617, "heldout": false, "domain": "reasoning", "ece": 0.1159557296037674, "proof_v2 (card)": NaN, "\u0394 vs proof_v2": NaN }, { "task": "yelp/score", "n": 500, "acc": 0.668, "chance": 0.2, "nll": 0.7834262755662202, "brier": 0.4423740240858843, "heldout": false, "domain": "classification", "ece": 0.09068749487400055, "proof_v2 (card)": NaN, "\u0394 vs proof_v2": NaN } ], "head_to_head": [ { "task": "ag_news/topic", "n": 120, "LightDec": 0.85, "proof_v2": 0.45, "\u0394": 0.39999999999999997 }, { "task": "agentharm/refuse", "n": 120, "LightDec": 0.6666666666666666, "proof_v2": 0.45, "\u0394": 0.21666666666666662 }, { "task": "agenttrek/finish_now", "n": 120, "LightDec": 0.7833333333333333, "proof_v2": 0.25833333333333336, "\u0394": 0.5249999999999999 }, { "task": "agenttrek/next_action_type", "n": 120, "LightDec": 0.875, "proof_v2": 0.075, "\u0394": 0.8 }, { "task": "anli/nli", "n": 120, "LightDec": 0.5916666666666667, "proof_v2": 0.325, "\u0394": 0.26666666666666666 }, { "task": "aqua_rat/math", "n": 120, "LightDec": 0.35, "proof_v2": 0.24166666666666667, "\u0394": 0.10833333333333331 }, { "task": "arc_challenge/mcq", "n": 120, "LightDec": 0.5333333333333333, "proof_v2": 0.30833333333333335, "\u0394": 0.22499999999999998 }, { "task": "arc_easy/mcq", "n": 120, "LightDec": 0.65, "proof_v2": 0.4583333333333333, "\u0394": 0.1916666666666667 }, { "task": "banking77/intent", "n": 120, "LightDec": 0.925, "proof_v2": 0.8666666666666667, "\u0394": 0.05833333333333335 }, { "task": "bigclonebench/clone", "n": 120, "LightDec": 0.9666666666666667, "proof_v2": 0.24166666666666667, "\u0394": 0.725 }, { "task": "bitext/category", "n": 120, "LightDec": 1.0, "proof_v2": 0.8583333333333333, "\u0394": 0.14166666666666672 }, { "task": "bitext/route", "n": 120, "LightDec": 1.0, "proof_v2": 0.925, "\u0394": 0.07499999999999996 }, { "task": "boolq/yes_no", "n": 120, "LightDec": 0.825, "proof_v2": 0.6666666666666666, "\u0394": 0.15833333333333333 }, { "task": "civil_comments/toxic", "n": 120, "LightDec": 0.9333333333333333, "proof_v2": 0.35833333333333334, "\u0394": 0.575 }, { "task": "clinc150/intent", "n": 120, "LightDec": 0.975, "proof_v2": 0.6416666666666667, "\u0394": 0.33333333333333326 }, { "task": "codexglue/code_to_doc", "n": 120, "LightDec": 1.0, "proof_v2": 0.9333333333333333, "\u0394": 0.06666666666666665 }, { "task": "codexglue/doc_to_code", "n": 120, "LightDec": 0.9833333333333333, "proof_v2": 0.975, "\u0394": 0.008333333333333304 }, { "task": "codexglue/func_name", "n": 120, "LightDec": 0.9916666666666667, "proof_v2": 0.9166666666666666, "\u0394": 0.07500000000000007 }, { "task": "codexglue/lang_id", "n": 120, "LightDec": 1.0, "proof_v2": 1.0, "\u0394": 0.0 }, { "task": "commonsense_qa/mcq", "n": 120, "LightDec": 0.6833333333333333, "proof_v2": 0.4083333333333333, "\u0394": 0.275 }, { "task": "counsel/critique_quality", "n": 120, "LightDec": 0.6083333333333333, "proof_v2": 0.25833333333333336, "\u0394": 0.3499999999999999 }, { "task": "counsel/step_has_error", "n": 120, "LightDec": 0.825, "proof_v2": 0.75, "\u0394": 0.07499999999999996 }, { "task": "devign/vulnerability", "n": 120, "LightDec": 0.6666666666666666, "proof_v2": 0.5666666666666667, "\u0394": 0.09999999999999998 }, { "task": "emotion/6way", "n": 120, "LightDec": 0.49166666666666664, "proof_v2": 0.36666666666666664, "\u0394": 0.125 }, { "task": "gsm8k/math", "n": 120, "LightDec": 0.7333333333333333, "proof_v2": 0.125, "\u0394": 0.6083333333333333 }, { "task": "hellaswag/continuation", "n": 120, "LightDec": 0.6083333333333333, "proof_v2": 0.2916666666666667, "\u0394": 0.3166666666666666 }, { "task": "hotpotqa/comparison_yes_no", "n": 26, "LightDec": 0.9230769230769231, "proof_v2": 0.46153846153846156, "\u0394": 0.46153846153846156 }, { "task": "hotpotqa/retrieve", "n": 120, "LightDec": 0.875, "proof_v2": 0.2916666666666667, "\u0394": 0.5833333333333333 }, { "task": "humaneval/completion", "n": 119, "LightDec": 0.8487394957983193, "proof_v2": 0.5126050420168067, "\u0394": 0.33613445378151263 }, { "task": "jailbreak/detect", "n": 120, "LightDec": 0.9833333333333333, "proof_v2": 0.5666666666666667, "\u0394": 0.41666666666666663 }, { "task": "long/contract_clause", "n": 120, "LightDec": 1.0, "proof_v2": 0.23333333333333334, "\u0394": 0.7666666666666666 }, { "task": "long/email_thread", "n": 120, "LightDec": 1.0, "proof_v2": 0.05, "\u0394": 0.95 }, { "task": "long/service_log", "n": 120, "LightDec": 1.0, "proof_v2": 0.24166666666666667, "\u0394": 0.7583333333333333 }, { "task": "massive_en/intent", "n": 120, "LightDec": 0.95, "proof_v2": 0.7416666666666667, "\u0394": 0.20833333333333326 }, { "task": "mbpp/bugspot", "n": 120, "LightDec": 0.8583333333333333, "proof_v2": 0.575, "\u0394": 0.2833333333333333 }, { "task": "mbpp/solution", "n": 120, "LightDec": 1.0, "proof_v2": 0.85, "\u0394": 0.15000000000000002 }, { "task": "mmlu/mcq", "n": 120, "LightDec": 0.4083333333333333, "proof_v2": 0.30833333333333335, "\u0394": 0.09999999999999998 }, { "task": "mnli/claim", "n": 120, "LightDec": 0.875, "proof_v2": 0.31666666666666665, "\u0394": 0.5583333333333333 }, { "task": "openbookqa/mcq", "n": 120, "LightDec": 0.65, "proof_v2": 0.2833333333333333, "\u0394": 0.3666666666666667 }, { "task": "policy/access_control_transfer", "n": 120, "LightDec": 1.0, "proof_v2": 0.23333333333333334, "\u0394": 0.7666666666666666 }, { "task": "policy/count_threshold_transfer", "n": 120, "LightDec": 0.8666666666666667, "proof_v2": 0.10833333333333334, "\u0394": 0.7583333333333333 }, { "task": "policy/free_shipping_transfer", "n": 120, "LightDec": 0.925, "proof_v2": 0.7083333333333334, "\u0394": 0.21666666666666667 }, { "task": "policy/invoice_overdue_transfer", "n": 120, "LightDec": 0.8083333333333333, "proof_v2": 0.6583333333333333, "\u0394": 0.15000000000000002 }, { "task": "policy/invoice_total_transfer", "n": 120, "LightDec": 0.45, "proof_v2": 0.525, "\u0394": -0.07500000000000001 }, { "task": "policy/refund_approval_transfer", "n": 120, "LightDec": 0.9583333333333334, "proof_v2": 0.325, "\u0394": 0.6333333333333333 }, { "task": "policy/return_window_transfer", "n": 120, "LightDec": 1.0, "proof_v2": 0.3416666666666667, "\u0394": 0.6583333333333333 }, { "task": "policy/sla_urgency_transfer", "n": 120, "LightDec": 0.5166666666666667, "proof_v2": 0.44166666666666665, "\u0394": 0.07500000000000007 }, { "task": "policy/table_compare_transfer", "n": 120, "LightDec": 0.9166666666666666, "proof_v2": 0.49166666666666664, "\u0394": 0.425 }, { "task": "policy/table_count_transfer", "n": 120, "LightDec": 0.31666666666666665, "proof_v2": 0.15833333333333333, "\u0394": 0.15833333333333333 }, { "task": "policy/table_extreme_transfer", "n": 120, "LightDec": 0.925, "proof_v2": 0.14166666666666666, "\u0394": 0.7833333333333334 }, { "task": "prompt_injections/detect", "n": 116, "LightDec": 0.5948275862068966, "proof_v2": 0.5431034482758621, "\u0394": 0.051724137931034475 }, { "task": "qasc/mcq", "n": 120, "LightDec": 0.9916666666666667, "proof_v2": 0.8583333333333333, "\u0394": 0.13333333333333341 }, { "task": "sciq/mcq", "n": 120, "LightDec": 0.9666666666666667, "proof_v2": 0.8666666666666667, "\u0394": 0.09999999999999998 }, { "task": "scitail/support", "n": 120, "LightDec": 0.9583333333333334, "proof_v2": 0.425, "\u0394": 0.5333333333333334 }, { "task": "snli/contradicts", "n": 120, "LightDec": 1.0, "proof_v2": 0.35833333333333334, "\u0394": 0.6416666666666666 }, { "task": "snli/must_be_true", "n": 120, "LightDec": 0.9916666666666667, "proof_v2": 0.925, "\u0394": 0.06666666666666665 }, { "task": "snli/nli", "n": 120, "LightDec": 0.9083333333333333, "proof_v2": 0.3416666666666667, "\u0394": 0.5666666666666667 }, { "task": "sst5/score", "n": 120, "LightDec": 0.38333333333333336, "proof_v2": 0.2833333333333333, "\u0394": 0.10000000000000003 }, { "task": "tev1_test/ag_news", "n": 120, "LightDec": 0.925, "proof_v2": 0.425, "\u0394": 0.5 }, { "task": "tev1_test/banking77", "n": 120, "LightDec": 0.7166666666666667, "proof_v2": 0.55, "\u0394": 0.16666666666666663 }, { "task": "tev1_test/boolq", "n": 120, "LightDec": 0.8416666666666667, "proof_v2": 0.5583333333333333, "\u0394": 0.2833333333333333 }, { "task": "tev1_test/mnli", "n": 120, "LightDec": 0.75, "proof_v2": 0.39166666666666666, "\u0394": 0.35833333333333334 }, { "task": "tev1_test/policy", "n": 120, "LightDec": 0.5, "proof_v2": 0.4083333333333333, "\u0394": 0.09166666666666667 }, { "task": "tev1_test/routing", "n": 120, "LightDec": 0.44166666666666665, "proof_v2": 0.2916666666666667, "\u0394": 0.14999999999999997 }, { "task": "tev1_test/sst5", "n": 120, "LightDec": 0.39166666666666666, "proof_v2": 0.275, "\u0394": 0.11666666666666664 }, { "task": "triage/support_email", "n": 120, "LightDec": 0.9583333333333334, "proof_v2": 0.39166666666666666, "\u0394": 0.5666666666666667 }, { "task": "typed_decisions/agent_trace_observability", "n": 120, "LightDec": 0.8083333333333333, "proof_v2": 0.3333333333333333, "\u0394": 0.47500000000000003 }, { "task": "typed_decisions/customer_service", "n": 120, "LightDec": 0.7416666666666667, "proof_v2": 0.2833333333333333, "\u0394": 0.45833333333333337 }, { "task": "typed_decisions/invoice_processing", "n": 120, "LightDec": 0.8166666666666667, "proof_v2": 0.475, "\u0394": 0.3416666666666667 }, { "task": "typed_decisions/security_incidents", "n": 120, "LightDec": 0.7583333333333333, "proof_v2": 0.5, "\u0394": 0.2583333333333333 }, { "task": "winogrande/blank", "n": 120, "LightDec": 0.6916666666666667, "proof_v2": 0.5583333333333333, "\u0394": 0.1333333333333333 }, { "task": "yelp/score", "n": 120, "LightDec": 0.625, "proof_v2": 0.325, "\u0394": 0.3 } ], "head_to_head_error": null, "v3_benchmarks": { "tev1": { "per_task": [ { "task": "tev1_test/ag_news", "n": 150, "acc": 0.92, "overlaps_LightDec_training_source": true }, { "task": "tev1_test/banking77", "n": 200, "acc": 0.72, "overlaps_LightDec_training_source": false }, { "task": "tev1_test/boolq", "n": 200, "acc": 0.835, "overlaps_LightDec_training_source": true }, { "task": "tev1_test/mnli", "n": 300, "acc": 0.7233333333333334, "overlaps_LightDec_training_source": true }, { "task": "tev1_test/policy", "n": 1200, "acc": 0.5291666666666667, "overlaps_LightDec_training_source": false }, { "task": "tev1_test/routing", "n": 600, "acc": 0.4583333333333333, "overlaps_LightDec_training_source": false }, { "task": "tev1_test/sst5", "n": 150, "acc": 0.3933333333333333, "overlaps_LightDec_training_source": false } ], "all": 0.5839285714285715, "no_overlap": 0.5176744186046511, "n": 2800, "published_tev1_4b": { "main decisions (1,000)": 0.88, "policy transfer (300)": 1.0 } }, "triage": { "emails": 101, "per_question": { "topic": 0.9900990099009901, "refund": 1.0, "breakage": 0.9900990099009901, "anger": 0.9207920792079208, "judgment": 0.8217821782178217 }, "pile_accuracy": 0.9405940594059405 }, "long_context": { "max_len": 2048, "per_task": [ { "task": "long/contract_clause", "n": 150, "acc": 1.0, "chance": 0.19999999999999996 }, { "task": "long/email_thread", "n": 150, "acc": 1.0, "chance": 0.25 }, { "task": "long/service_log", "n": 150, "acc": 1.0, "chance": 0.19999999999999996 } ] } }, "architecture": "FalconDec", "latency": { "gpu": "NVIDIA RTX PRO 6000 Blackwell Server Edition", "calls": [ { "device": "cuda", "weights": "fp16", "questions": 1, "p50_ms": 10.05, "p95_ms": 10.43 }, { "device": "cuda", "weights": "fp16", "questions": 5, "p50_ms": 11.32, "p95_ms": 11.41 }, { "device": "cuda", "weights": "fp16", "questions": 10, "p50_ms": 12.57, "p95_ms": 12.69 }, { "device": "cpu fp32", "weights": "int8 file", "questions": 1, "p50_ms": 48.4, "p95_ms": 48.8 } ], "throughput_per_s": 2588.8502560238694 }, "env": { "torch": "2.9.0+cu130", "transformers": "5.17.0", "python": "3.12.12", "gpu": "NVIDIA RTX PRO 6000 Blackwell Server Edition" } }