Upload 3 files
Browse files- config.json +40 -0
- model.safetensors +3 -0
- results_full.json +670 -0
config.json
ADDED
|
@@ -0,0 +1,40 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"tier": "base",
|
| 3 |
+
"d": 512,
|
| 4 |
+
"heads": 8,
|
| 5 |
+
"e": 128,
|
| 6 |
+
"buckets": 65536,
|
| 7 |
+
"recursions": 6,
|
| 8 |
+
"interact": 2,
|
| 9 |
+
"mlp": 1536,
|
| 10 |
+
"lr": 0.00015,
|
| 11 |
+
"budget": 16384,
|
| 12 |
+
"layout": {
|
| 13 |
+
"max_len": 1024,
|
| 14 |
+
"long_max_len": 2048,
|
| 15 |
+
"long_opts_threshold": 24,
|
| 16 |
+
"head_max_len": 192,
|
| 17 |
+
"max_tok_per_opt": 24
|
| 18 |
+
},
|
| 19 |
+
"temperatures": [
|
| 20 |
+
[
|
| 21 |
+
2.223009569218283,
|
| 22 |
+
3.850024543056701,
|
| 23 |
+
4.047131644703235,
|
| 24 |
+
1.5673085376630793
|
| 25 |
+
],
|
| 26 |
+
[
|
| 27 |
+
7.009216863860871,
|
| 28 |
+
7.009216863860871,
|
| 29 |
+
7.009216863860871,
|
| 30 |
+
7.009216863860871
|
| 31 |
+
],
|
| 32 |
+
[
|
| 33 |
+
11.548028041677576,
|
| 34 |
+
11.548028041677576,
|
| 35 |
+
11.548028041677576,
|
| 36 |
+
11.548028041677576
|
| 37 |
+
]
|
| 38 |
+
],
|
| 39 |
+
"pretrain": null
|
| 40 |
+
}
|
model.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:2b37b2dc7c7021b8f359cffbcb43a3784618b59d4b7ad14f153e5149ec0dd740
|
| 3 |
+
size 24537642
|
results_full.json
ADDED
|
@@ -0,0 +1,670 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"tiers": {
|
| 3 |
+
"base": {
|
| 4 |
+
"macro": 0.5841957348300755,
|
| 5 |
+
"micro": 0.5956310862127308,
|
| 6 |
+
"per_domain": {
|
| 7 |
+
"agentic": 0.6046159274822298,
|
| 8 |
+
"classification": 0.4305,
|
| 9 |
+
"code": 0.7406625660653666,
|
| 10 |
+
"guardrails": 0.690390802488509,
|
| 11 |
+
"intents": 0.5826666666666667,
|
| 12 |
+
"policy": 0.5653333333333335,
|
| 13 |
+
"reasoning": 0.4717053299448031,
|
| 14 |
+
"support": 0.995,
|
| 15 |
+
"workflows": 0.604
|
| 16 |
+
},
|
| 17 |
+
"per_task": {
|
| 18 |
+
"arc_easy/mcq": 0.304,
|
| 19 |
+
"arc_challenge/mcq": 0.294,
|
| 20 |
+
"openbookqa/mcq": 0.362,
|
| 21 |
+
"sciq/mcq": 0.608433734939759,
|
| 22 |
+
"commonsense_qa/mcq": 0.19066937119675456,
|
| 23 |
+
"qasc/mcq": 0.468,
|
| 24 |
+
"hellaswag/continuation": 0.28,
|
| 25 |
+
"winogrande/blank": 0.562,
|
| 26 |
+
"mmlu/mcq": 0.284,
|
| 27 |
+
"boolq/yes_no": 0.626,
|
| 28 |
+
"gsm8k/math": 0.306,
|
| 29 |
+
"aqua_rat/math": 0.21862348178137653,
|
| 30 |
+
"snli/nli": 0.7236842105263158,
|
| 31 |
+
"snli/must_be_true": 0.908,
|
| 32 |
+
"snli/contradicts": 0.854,
|
| 33 |
+
"mnli/claim": 0.52,
|
| 34 |
+
"anli/nli": 0.321285140562249,
|
| 35 |
+
"scitail/support": 0.66,
|
| 36 |
+
"clinc150/intent": 0.848,
|
| 37 |
+
"banking77/intent": 0.582,
|
| 38 |
+
"banking77/intent_77": 0.04666666666666667,
|
| 39 |
+
"massive_en/intent": 0.854,
|
| 40 |
+
"bitext/route": 0.994,
|
| 41 |
+
"bitext/category": 0.996,
|
| 42 |
+
"ag_news/topic": 0.846,
|
| 43 |
+
"yelp/score": 0.486,
|
| 44 |
+
"emotion/6way": 0.116,
|
| 45 |
+
"sst5/score": 0.274,
|
| 46 |
+
"jailbreak/detect": 0.9122137404580153,
|
| 47 |
+
"prompt_injections/detect": 0.47413793103448276,
|
| 48 |
+
"civil_comments/toxic": 0.868,
|
| 49 |
+
"codexglue/lang_id": 0.998015873015873,
|
| 50 |
+
"codexglue/code_to_doc": 0.8392857142857143,
|
| 51 |
+
"codexglue/doc_to_code": 0.8849206349206349,
|
| 52 |
+
"codexglue/func_name": 0.6980728051391863,
|
| 53 |
+
"devign/vulnerability": 0.642,
|
| 54 |
+
"bigclonebench/clone": 0.954,
|
| 55 |
+
"mbpp/solution": 0.792,
|
| 56 |
+
"mbpp/bugspot": 0.4375,
|
| 57 |
+
"humaneval/completion": 0.42016806722689076,
|
| 58 |
+
"policy/return_window_transfer": 0.826,
|
| 59 |
+
"policy/refund_approval_transfer": 0.52,
|
| 60 |
+
"policy/free_shipping_transfer": 0.8,
|
| 61 |
+
"policy/invoice_overdue_transfer": 0.488,
|
| 62 |
+
"policy/table_extreme_transfer": 0.284,
|
| 63 |
+
"policy/count_threshold_transfer": 0.348,
|
| 64 |
+
"policy/sla_urgency_transfer": 0.302,
|
| 65 |
+
"policy/access_control_transfer": 1.0,
|
| 66 |
+
"policy/invoice_total_transfer": 0.52,
|
| 67 |
+
"typed_decisions/security_incidents": 0.658,
|
| 68 |
+
"typed_decisions/agent_trace_observability": 0.698,
|
| 69 |
+
"typed_decisions/invoice_processing": 0.56,
|
| 70 |
+
"typed_decisions/customer_service": 0.5,
|
| 71 |
+
"counsel/critique_quality": 0.4577114427860697,
|
| 72 |
+
"counsel/step_has_error": 0.7213930348258707,
|
| 73 |
+
"hotpotqa/retrieve": 0.5855130784708249,
|
| 74 |
+
"hotpotqa/comparison_yes_no": 0.6538461538461539,
|
| 75 |
+
"agentharm/refuse": 0.5072115384615384
|
| 76 |
+
},
|
| 77 |
+
"bytes": 24538267,
|
| 78 |
+
"ece": {
|
| 79 |
+
"before": 0.2459549366365098,
|
| 80 |
+
"after": 0.029557739478093977
|
| 81 |
+
},
|
| 82 |
+
"speed": {
|
| 83 |
+
"device": "cuda (NVIDIA RTX PRO 6000 Blackwell Server Edition)",
|
| 84 |
+
"n": 200,
|
| 85 |
+
"median_ms": 5.14031700004125,
|
| 86 |
+
"p90_ms": 5.255812798714032,
|
| 87 |
+
"mean_ms": 5.145381490110594,
|
| 88 |
+
"batch128_per_s": 7475.691122289575
|
| 89 |
+
},
|
| 90 |
+
"agentic": {
|
| 91 |
+
"top_probability": {
|
| 92 |
+
"overall": {
|
| 93 |
+
"n": 26597,
|
| 94 |
+
"acted": 5713,
|
| 95 |
+
"right_when_acts": 0.8974269210572379,
|
| 96 |
+
"flags_mistakes": 0.9455137145513715,
|
| 97 |
+
"accuracy": 0.5956310862127308,
|
| 98 |
+
"handles_alone": 0.21479866150317706
|
| 99 |
+
},
|
| 100 |
+
"per_domain": {
|
| 101 |
+
"agentic": {
|
| 102 |
+
"n": 925,
|
| 103 |
+
"acted": 165,
|
| 104 |
+
"right_when_acts": 0.8,
|
| 105 |
+
"flags_mistakes": 0.9131578947368421,
|
| 106 |
+
"accuracy": 0.5891891891891892,
|
| 107 |
+
"handles_alone": 0.1783783783783784
|
| 108 |
+
},
|
| 109 |
+
"classification": {
|
| 110 |
+
"n": 2000,
|
| 111 |
+
"acted": 137,
|
| 112 |
+
"right_when_acts": 0.9927007299270073,
|
| 113 |
+
"flags_mistakes": 0.9991220368744512,
|
| 114 |
+
"accuracy": 0.4305,
|
| 115 |
+
"handles_alone": 0.0685
|
| 116 |
+
},
|
| 117 |
+
"code": {
|
| 118 |
+
"n": 3854,
|
| 119 |
+
"acted": 1330,
|
| 120 |
+
"right_when_acts": 0.9887218045112782,
|
| 121 |
+
"flags_mistakes": 0.98125,
|
| 122 |
+
"accuracy": 0.7924234561494551,
|
| 123 |
+
"handles_alone": 0.34509600415153086
|
| 124 |
+
},
|
| 125 |
+
"guardrails": {
|
| 126 |
+
"n": 1294,
|
| 127 |
+
"acted": 209,
|
| 128 |
+
"right_when_acts": 0.937799043062201,
|
| 129 |
+
"flags_mistakes": 0.9633802816901409,
|
| 130 |
+
"accuracy": 0.7256568778979907,
|
| 131 |
+
"handles_alone": 0.1615146831530139
|
| 132 |
+
},
|
| 133 |
+
"intents": {
|
| 134 |
+
"n": 1800,
|
| 135 |
+
"acted": 697,
|
| 136 |
+
"right_when_acts": 0.9096126255380201,
|
| 137 |
+
"flags_mistakes": 0.9021739130434783,
|
| 138 |
+
"accuracy": 0.6422222222222222,
|
| 139 |
+
"handles_alone": 0.38722222222222225
|
| 140 |
+
},
|
| 141 |
+
"policy": {
|
| 142 |
+
"n": 4500,
|
| 143 |
+
"acted": 1854,
|
| 144 |
+
"right_when_acts": 0.7740021574973032,
|
| 145 |
+
"flags_mistakes": 0.7857873210633947,
|
| 146 |
+
"accuracy": 0.5653333333333334,
|
| 147 |
+
"handles_alone": 0.412
|
| 148 |
+
},
|
| 149 |
+
"reasoning": {
|
| 150 |
+
"n": 9224,
|
| 151 |
+
"acted": 351,
|
| 152 |
+
"right_when_acts": 0.8803418803418803,
|
| 153 |
+
"flags_mistakes": 0.9910333048676345,
|
| 154 |
+
"accuracy": 0.492194275802255,
|
| 155 |
+
"handles_alone": 0.038052905464006935
|
| 156 |
+
},
|
| 157 |
+
"support": {
|
| 158 |
+
"n": 1000,
|
| 159 |
+
"acted": 970,
|
| 160 |
+
"right_when_acts": 1.0,
|
| 161 |
+
"flags_mistakes": 1.0,
|
| 162 |
+
"accuracy": 0.995,
|
| 163 |
+
"handles_alone": 0.97
|
| 164 |
+
},
|
| 165 |
+
"workflows": {
|
| 166 |
+
"n": 2000,
|
| 167 |
+
"acted": 0,
|
| 168 |
+
"right_when_acts": NaN,
|
| 169 |
+
"flags_mistakes": 1.0,
|
| 170 |
+
"accuracy": 0.604,
|
| 171 |
+
"handles_alone": 0.0
|
| 172 |
+
}
|
| 173 |
+
}
|
| 174 |
+
},
|
| 175 |
+
"laya": {
|
| 176 |
+
"overall": {
|
| 177 |
+
"n": 26597,
|
| 178 |
+
"acted": 4101,
|
| 179 |
+
"right_when_acts": 0.9222140941233845,
|
| 180 |
+
"flags_mistakes": 0.9703393770339377,
|
| 181 |
+
"accuracy": 0.5956310862127308,
|
| 182 |
+
"handles_alone": 0.15419032221679138
|
| 183 |
+
},
|
| 184 |
+
"per_domain": {
|
| 185 |
+
"agentic": {
|
| 186 |
+
"n": 925,
|
| 187 |
+
"acted": 152,
|
| 188 |
+
"right_when_acts": 0.7960526315789473,
|
| 189 |
+
"flags_mistakes": 0.9184210526315789,
|
| 190 |
+
"accuracy": 0.5891891891891892,
|
| 191 |
+
"handles_alone": 0.1643243243243243
|
| 192 |
+
},
|
| 193 |
+
"classification": {
|
| 194 |
+
"n": 2000,
|
| 195 |
+
"acted": 9,
|
| 196 |
+
"right_when_acts": 1.0,
|
| 197 |
+
"flags_mistakes": 1.0,
|
| 198 |
+
"accuracy": 0.4305,
|
| 199 |
+
"handles_alone": 0.0045
|
| 200 |
+
},
|
| 201 |
+
"code": {
|
| 202 |
+
"n": 3854,
|
| 203 |
+
"acted": 978,
|
| 204 |
+
"right_when_acts": 0.9887525562372188,
|
| 205 |
+
"flags_mistakes": 0.98625,
|
| 206 |
+
"accuracy": 0.7924234561494551,
|
| 207 |
+
"handles_alone": 0.2537623248572911
|
| 208 |
+
},
|
| 209 |
+
"guardrails": {
|
| 210 |
+
"n": 1294,
|
| 211 |
+
"acted": 209,
|
| 212 |
+
"right_when_acts": 0.937799043062201,
|
| 213 |
+
"flags_mistakes": 0.9633802816901409,
|
| 214 |
+
"accuracy": 0.7256568778979907,
|
| 215 |
+
"handles_alone": 0.1615146831530139
|
| 216 |
+
},
|
| 217 |
+
"intents": {
|
| 218 |
+
"n": 1800,
|
| 219 |
+
"acted": 428,
|
| 220 |
+
"right_when_acts": 0.9392523364485982,
|
| 221 |
+
"flags_mistakes": 0.9596273291925466,
|
| 222 |
+
"accuracy": 0.6422222222222222,
|
| 223 |
+
"handles_alone": 0.23777777777777778
|
| 224 |
+
},
|
| 225 |
+
"policy": {
|
| 226 |
+
"n": 4500,
|
| 227 |
+
"acted": 1340,
|
| 228 |
+
"right_when_acts": 0.8373134328358209,
|
| 229 |
+
"flags_mistakes": 0.8885480572597138,
|
| 230 |
+
"accuracy": 0.5653333333333334,
|
| 231 |
+
"handles_alone": 0.29777777777777775
|
| 232 |
+
},
|
| 233 |
+
"reasoning": {
|
| 234 |
+
"n": 9224,
|
| 235 |
+
"acted": 77,
|
| 236 |
+
"right_when_acts": 0.7402597402597403,
|
| 237 |
+
"flags_mistakes": 0.9957301451750641,
|
| 238 |
+
"accuracy": 0.492194275802255,
|
| 239 |
+
"handles_alone": 0.008347788378143973
|
| 240 |
+
},
|
| 241 |
+
"support": {
|
| 242 |
+
"n": 1000,
|
| 243 |
+
"acted": 908,
|
| 244 |
+
"right_when_acts": 1.0,
|
| 245 |
+
"flags_mistakes": 1.0,
|
| 246 |
+
"accuracy": 0.995,
|
| 247 |
+
"handles_alone": 0.908
|
| 248 |
+
},
|
| 249 |
+
"workflows": {
|
| 250 |
+
"n": 2000,
|
| 251 |
+
"acted": 0,
|
| 252 |
+
"right_when_acts": NaN,
|
| 253 |
+
"flags_mistakes": 1.0,
|
| 254 |
+
"accuracy": 0.604,
|
| 255 |
+
"handles_alone": 0.0
|
| 256 |
+
}
|
| 257 |
+
}
|
| 258 |
+
}
|
| 259 |
+
}
|
| 260 |
+
}
|
| 261 |
+
},
|
| 262 |
+
"tasks": {
|
| 263 |
+
"arc_easy/mcq": {
|
| 264 |
+
"domain": "reasoning",
|
| 265 |
+
"test_only": true,
|
| 266 |
+
"n": 500,
|
| 267 |
+
"chance": 0.24980000000000002,
|
| 268 |
+
"majority": 0.008
|
| 269 |
+
},
|
| 270 |
+
"arc_challenge/mcq": {
|
| 271 |
+
"domain": "reasoning",
|
| 272 |
+
"test_only": true,
|
| 273 |
+
"n": 500,
|
| 274 |
+
"chance": 0.25006666666666666,
|
| 275 |
+
"majority": 0.004
|
| 276 |
+
},
|
| 277 |
+
"openbookqa/mcq": {
|
| 278 |
+
"domain": "reasoning",
|
| 279 |
+
"test_only": false,
|
| 280 |
+
"n": 500,
|
| 281 |
+
"chance": 0.25,
|
| 282 |
+
"majority": 0.006
|
| 283 |
+
},
|
| 284 |
+
"sciq/mcq": {
|
| 285 |
+
"domain": "reasoning",
|
| 286 |
+
"test_only": false,
|
| 287 |
+
"n": 498,
|
| 288 |
+
"chance": 0.25,
|
| 289 |
+
"majority": 0.012048192771084338
|
| 290 |
+
},
|
| 291 |
+
"commonsense_qa/mcq": {
|
| 292 |
+
"domain": "reasoning",
|
| 293 |
+
"test_only": false,
|
| 294 |
+
"n": 493,
|
| 295 |
+
"chance": 0.20000000000000176,
|
| 296 |
+
"majority": 0.008113590263691683
|
| 297 |
+
},
|
| 298 |
+
"qasc/mcq": {
|
| 299 |
+
"domain": "reasoning",
|
| 300 |
+
"test_only": false,
|
| 301 |
+
"n": 500,
|
| 302 |
+
"chance": 0.125,
|
| 303 |
+
"majority": 0.01
|
| 304 |
+
},
|
| 305 |
+
"hellaswag/continuation": {
|
| 306 |
+
"domain": "reasoning",
|
| 307 |
+
"test_only": false,
|
| 308 |
+
"n": 500,
|
| 309 |
+
"chance": 0.25,
|
| 310 |
+
"majority": 0.002
|
| 311 |
+
},
|
| 312 |
+
"winogrande/blank": {
|
| 313 |
+
"domain": "reasoning",
|
| 314 |
+
"test_only": false,
|
| 315 |
+
"n": 500,
|
| 316 |
+
"chance": 0.5,
|
| 317 |
+
"majority": 0.024
|
| 318 |
+
},
|
| 319 |
+
"mmlu/mcq": {
|
| 320 |
+
"domain": "reasoning",
|
| 321 |
+
"test_only": true,
|
| 322 |
+
"n": 500,
|
| 323 |
+
"chance": 0.25,
|
| 324 |
+
"majority": 0.022
|
| 325 |
+
},
|
| 326 |
+
"boolq/yes_no": {
|
| 327 |
+
"domain": "reasoning",
|
| 328 |
+
"test_only": false,
|
| 329 |
+
"n": 500,
|
| 330 |
+
"chance": 0.5,
|
| 331 |
+
"majority": 0.58
|
| 332 |
+
},
|
| 333 |
+
"gsm8k/math": {
|
| 334 |
+
"domain": "reasoning",
|
| 335 |
+
"test_only": false,
|
| 336 |
+
"n": 500,
|
| 337 |
+
"chance": 0.25,
|
| 338 |
+
"majority": 0.034
|
| 339 |
+
},
|
| 340 |
+
"aqua_rat/math": {
|
| 341 |
+
"domain": "reasoning",
|
| 342 |
+
"test_only": false,
|
| 343 |
+
"n": 247,
|
| 344 |
+
"chance": 0.20000000000000065,
|
| 345 |
+
"majority": 0.020242914979757085
|
| 346 |
+
},
|
| 347 |
+
"snli/nli": {
|
| 348 |
+
"domain": "reasoning",
|
| 349 |
+
"test_only": false,
|
| 350 |
+
"n": 988,
|
| 351 |
+
"chance": 0.33333333333333204,
|
| 352 |
+
"majority": 0.3431174089068826
|
| 353 |
+
},
|
| 354 |
+
"snli/must_be_true": {
|
| 355 |
+
"domain": "reasoning",
|
| 356 |
+
"test_only": false,
|
| 357 |
+
"n": 500,
|
| 358 |
+
"chance": 0.33333333333333404,
|
| 359 |
+
"majority": 0.004
|
| 360 |
+
},
|
| 361 |
+
"snli/contradicts": {
|
| 362 |
+
"domain": "reasoning",
|
| 363 |
+
"test_only": false,
|
| 364 |
+
"n": 500,
|
| 365 |
+
"chance": 0.33333333333333404,
|
| 366 |
+
"majority": 0.004
|
| 367 |
+
},
|
| 368 |
+
"mnli/claim": {
|
| 369 |
+
"domain": "reasoning",
|
| 370 |
+
"test_only": false,
|
| 371 |
+
"n": 500,
|
| 372 |
+
"chance": 0.33333333333333404,
|
| 373 |
+
"majority": 0.352
|
| 374 |
+
},
|
| 375 |
+
"anli/nli": {
|
| 376 |
+
"domain": "reasoning",
|
| 377 |
+
"test_only": false,
|
| 378 |
+
"n": 498,
|
| 379 |
+
"chance": 0.33333333333333404,
|
| 380 |
+
"majority": 0.3534136546184739
|
| 381 |
+
},
|
| 382 |
+
"scitail/support": {
|
| 383 |
+
"domain": "reasoning",
|
| 384 |
+
"test_only": false,
|
| 385 |
+
"n": 500,
|
| 386 |
+
"chance": 0.5,
|
| 387 |
+
"majority": 0.618
|
| 388 |
+
},
|
| 389 |
+
"clinc150/intent": {
|
| 390 |
+
"domain": "intents",
|
| 391 |
+
"test_only": false,
|
| 392 |
+
"n": 500,
|
| 393 |
+
"chance": 0.12404474984829378,
|
| 394 |
+
"majority": 0.184
|
| 395 |
+
},
|
| 396 |
+
"banking77/intent": {
|
| 397 |
+
"domain": "intents",
|
| 398 |
+
"test_only": true,
|
| 399 |
+
"n": 500,
|
| 400 |
+
"chance": 0.3214666666666664,
|
| 401 |
+
"majority": 0.026
|
| 402 |
+
},
|
| 403 |
+
"banking77/intent_77": {
|
| 404 |
+
"domain": "intents",
|
| 405 |
+
"test_only": true,
|
| 406 |
+
"n": 300,
|
| 407 |
+
"chance": 0.012987012987013017,
|
| 408 |
+
"majority": 0.03
|
| 409 |
+
},
|
| 410 |
+
"massive_en/intent": {
|
| 411 |
+
"domain": "intents",
|
| 412 |
+
"test_only": false,
|
| 413 |
+
"n": 500,
|
| 414 |
+
"chance": 0.12958727789085667,
|
| 415 |
+
"majority": 0.06
|
| 416 |
+
},
|
| 417 |
+
"bitext/route": {
|
| 418 |
+
"domain": "support",
|
| 419 |
+
"test_only": false,
|
| 420 |
+
"n": 500,
|
| 421 |
+
"chance": 0.19016313131313176,
|
| 422 |
+
"majority": 0.102
|
| 423 |
+
},
|
| 424 |
+
"bitext/category": {
|
| 425 |
+
"domain": "support",
|
| 426 |
+
"test_only": false,
|
| 427 |
+
"n": 500,
|
| 428 |
+
"chance": 0.1612812409812413,
|
| 429 |
+
"majority": 0.174
|
| 430 |
+
},
|
| 431 |
+
"ag_news/topic": {
|
| 432 |
+
"domain": "classification",
|
| 433 |
+
"test_only": false,
|
| 434 |
+
"n": 500,
|
| 435 |
+
"chance": 0.25,
|
| 436 |
+
"majority": 0.136
|
| 437 |
+
},
|
| 438 |
+
"yelp/score": {
|
| 439 |
+
"domain": "classification",
|
| 440 |
+
"test_only": false,
|
| 441 |
+
"n": 500,
|
| 442 |
+
"chance": 0.20000000000000176,
|
| 443 |
+
"majority": 0.098
|
| 444 |
+
},
|
| 445 |
+
"emotion/6way": {
|
| 446 |
+
"domain": "classification",
|
| 447 |
+
"test_only": true,
|
| 448 |
+
"n": 500,
|
| 449 |
+
"chance": 0.16666666666666702,
|
| 450 |
+
"majority": 0.386
|
| 451 |
+
},
|
| 452 |
+
"sst5/score": {
|
| 453 |
+
"domain": "classification",
|
| 454 |
+
"test_only": true,
|
| 455 |
+
"n": 500,
|
| 456 |
+
"chance": 0.20000000000000176,
|
| 457 |
+
"majority": 0.27
|
| 458 |
+
},
|
| 459 |
+
"jailbreak/detect": {
|
| 460 |
+
"domain": "guardrails",
|
| 461 |
+
"test_only": false,
|
| 462 |
+
"n": 262,
|
| 463 |
+
"chance": 0.5,
|
| 464 |
+
"majority": 0.5305343511450382
|
| 465 |
+
},
|
| 466 |
+
"prompt_injections/detect": {
|
| 467 |
+
"domain": "guardrails",
|
| 468 |
+
"test_only": true,
|
| 469 |
+
"n": 116,
|
| 470 |
+
"chance": 0.5,
|
| 471 |
+
"majority": 0.5172413793103449
|
| 472 |
+
},
|
| 473 |
+
"civil_comments/toxic": {
|
| 474 |
+
"domain": "guardrails",
|
| 475 |
+
"test_only": false,
|
| 476 |
+
"n": 500,
|
| 477 |
+
"chance": 0.5,
|
| 478 |
+
"majority": 0.5
|
| 479 |
+
},
|
| 480 |
+
"codexglue/lang_id": {
|
| 481 |
+
"domain": "code",
|
| 482 |
+
"test_only": false,
|
| 483 |
+
"n": 504,
|
| 484 |
+
"chance": 0.23541666666666738,
|
| 485 |
+
"majority": 0.16666666666666666
|
| 486 |
+
},
|
| 487 |
+
"codexglue/code_to_doc": {
|
| 488 |
+
"domain": "code",
|
| 489 |
+
"test_only": false,
|
| 490 |
+
"n": 504,
|
| 491 |
+
"chance": 0.2623346560846559,
|
| 492 |
+
"majority": 0.001984126984126984
|
| 493 |
+
},
|
| 494 |
+
"codexglue/doc_to_code": {
|
| 495 |
+
"domain": "code",
|
| 496 |
+
"test_only": false,
|
| 497 |
+
"n": 504,
|
| 498 |
+
"chance": 0.278604497354497,
|
| 499 |
+
"majority": 0.001984126984126984
|
| 500 |
+
},
|
| 501 |
+
"codexglue/func_name": {
|
| 502 |
+
"domain": "code",
|
| 503 |
+
"test_only": false,
|
| 504 |
+
"n": 467,
|
| 505 |
+
"chance": 0.25499643112062836,
|
| 506 |
+
"majority": 0.006423982869379015
|
| 507 |
+
},
|
| 508 |
+
"devign/vulnerability": {
|
| 509 |
+
"domain": "code",
|
| 510 |
+
"test_only": false,
|
| 511 |
+
"n": 500,
|
| 512 |
+
"chance": 0.5,
|
| 513 |
+
"majority": 0.574
|
| 514 |
+
},
|
| 515 |
+
"bigclonebench/clone": {
|
| 516 |
+
"domain": "code",
|
| 517 |
+
"test_only": false,
|
| 518 |
+
"n": 500,
|
| 519 |
+
"chance": 0.5,
|
| 520 |
+
"majority": 0.852
|
| 521 |
+
},
|
| 522 |
+
"mbpp/solution": {
|
| 523 |
+
"domain": "code",
|
| 524 |
+
"test_only": false,
|
| 525 |
+
"n": 500,
|
| 526 |
+
"chance": 0.25,
|
| 527 |
+
"majority": 0.004
|
| 528 |
+
},
|
| 529 |
+
"mbpp/bugspot": {
|
| 530 |
+
"domain": "code",
|
| 531 |
+
"test_only": false,
|
| 532 |
+
"n": 256,
|
| 533 |
+
"chance": 0.4055989583333324,
|
| 534 |
+
"majority": 0.00390625
|
| 535 |
+
},
|
| 536 |
+
"humaneval/completion": {
|
| 537 |
+
"domain": "code",
|
| 538 |
+
"test_only": true,
|
| 539 |
+
"n": 119,
|
| 540 |
+
"chance": 0.39355742296918794,
|
| 541 |
+
"majority": 0.008403361344537815
|
| 542 |
+
},
|
| 543 |
+
"policy/return_window_transfer": {
|
| 544 |
+
"domain": "policy",
|
| 545 |
+
"test_only": true,
|
| 546 |
+
"n": 500,
|
| 547 |
+
"chance": 0.33333333333333404,
|
| 548 |
+
"majority": 0.422
|
| 549 |
+
},
|
| 550 |
+
"policy/refund_approval_transfer": {
|
| 551 |
+
"domain": "policy",
|
| 552 |
+
"test_only": true,
|
| 553 |
+
"n": 500,
|
| 554 |
+
"chance": 0.33333333333333404,
|
| 555 |
+
"majority": 0.522
|
| 556 |
+
},
|
| 557 |
+
"policy/free_shipping_transfer": {
|
| 558 |
+
"domain": "policy",
|
| 559 |
+
"test_only": true,
|
| 560 |
+
"n": 500,
|
| 561 |
+
"chance": 0.5,
|
| 562 |
+
"majority": 0.534
|
| 563 |
+
},
|
| 564 |
+
"policy/invoice_overdue_transfer": {
|
| 565 |
+
"domain": "policy",
|
| 566 |
+
"test_only": true,
|
| 567 |
+
"n": 500,
|
| 568 |
+
"chance": 0.5,
|
| 569 |
+
"majority": 0.616
|
| 570 |
+
},
|
| 571 |
+
"policy/table_extreme_transfer": {
|
| 572 |
+
"domain": "policy",
|
| 573 |
+
"test_only": true,
|
| 574 |
+
"n": 500,
|
| 575 |
+
"chance": 0.23716666666666722,
|
| 576 |
+
"majority": 0.056
|
| 577 |
+
},
|
| 578 |
+
"policy/count_threshold_transfer": {
|
| 579 |
+
"domain": "policy",
|
| 580 |
+
"test_only": true,
|
| 581 |
+
"n": 500,
|
| 582 |
+
"chance": 0.17848809523809558,
|
| 583 |
+
"majority": 0.23
|
| 584 |
+
},
|
| 585 |
+
"policy/sla_urgency_transfer": {
|
| 586 |
+
"domain": "policy",
|
| 587 |
+
"test_only": true,
|
| 588 |
+
"n": 500,
|
| 589 |
+
"chance": 0.25,
|
| 590 |
+
"majority": 0.402
|
| 591 |
+
},
|
| 592 |
+
"policy/access_control_transfer": {
|
| 593 |
+
"domain": "policy",
|
| 594 |
+
"test_only": true,
|
| 595 |
+
"n": 500,
|
| 596 |
+
"chance": 0.33333333333333404,
|
| 597 |
+
"majority": 0.626
|
| 598 |
+
},
|
| 599 |
+
"policy/invoice_total_transfer": {
|
| 600 |
+
"domain": "policy",
|
| 601 |
+
"test_only": true,
|
| 602 |
+
"n": 500,
|
| 603 |
+
"chance": 0.5,
|
| 604 |
+
"majority": 0.54
|
| 605 |
+
},
|
| 606 |
+
"typed_decisions/security_incidents": {
|
| 607 |
+
"domain": "workflows",
|
| 608 |
+
"test_only": false,
|
| 609 |
+
"n": 500,
|
| 610 |
+
"chance": 0.33999999999999975,
|
| 611 |
+
"majority": 0.148
|
| 612 |
+
},
|
| 613 |
+
"typed_decisions/agent_trace_observability": {
|
| 614 |
+
"domain": "workflows",
|
| 615 |
+
"test_only": false,
|
| 616 |
+
"n": 500,
|
| 617 |
+
"chance": 0.3,
|
| 618 |
+
"majority": 0.11
|
| 619 |
+
},
|
| 620 |
+
"typed_decisions/invoice_processing": {
|
| 621 |
+
"domain": "workflows",
|
| 622 |
+
"test_only": false,
|
| 623 |
+
"n": 500,
|
| 624 |
+
"chance": 0.35,
|
| 625 |
+
"majority": 0.178
|
| 626 |
+
},
|
| 627 |
+
"typed_decisions/customer_service": {
|
| 628 |
+
"domain": "workflows",
|
| 629 |
+
"test_only": false,
|
| 630 |
+
"n": 500,
|
| 631 |
+
"chance": 0.28000000000000036,
|
| 632 |
+
"majority": 0.134
|
| 633 |
+
},
|
| 634 |
+
"counsel/critique_quality": {
|
| 635 |
+
"domain": "agentic",
|
| 636 |
+
"test_only": false,
|
| 637 |
+
"n": 201,
|
| 638 |
+
"chance": 0.333333333333334,
|
| 639 |
+
"majority": 0.48258706467661694
|
| 640 |
+
},
|
| 641 |
+
"counsel/step_has_error": {
|
| 642 |
+
"domain": "agentic",
|
| 643 |
+
"test_only": false,
|
| 644 |
+
"n": 201,
|
| 645 |
+
"chance": 0.5,
|
| 646 |
+
"majority": 0.7562189054726368
|
| 647 |
+
},
|
| 648 |
+
"hotpotqa/retrieve": {
|
| 649 |
+
"domain": "agentic",
|
| 650 |
+
"test_only": false,
|
| 651 |
+
"n": 497,
|
| 652 |
+
"chance": 0.1674172016224337,
|
| 653 |
+
"majority": 0.004024144869215292
|
| 654 |
+
},
|
| 655 |
+
"hotpotqa/comparison_yes_no": {
|
| 656 |
+
"domain": "agentic",
|
| 657 |
+
"test_only": false,
|
| 658 |
+
"n": 26,
|
| 659 |
+
"chance": 0.5,
|
| 660 |
+
"majority": 0.5384615384615384
|
| 661 |
+
},
|
| 662 |
+
"agentharm/refuse": {
|
| 663 |
+
"domain": "guardrails",
|
| 664 |
+
"test_only": true,
|
| 665 |
+
"n": 416,
|
| 666 |
+
"chance": 0.5,
|
| 667 |
+
"majority": 0.5
|
| 668 |
+
}
|
| 669 |
+
}
|
| 670 |
+
}
|