PEFT
Safetensors
English
sev
research
cybersecurity
agent-activity
decision-model
lora
macmacmacmac commited on
Commit
7f9f3f5
·
verified ·
1 Parent(s): da0a131

Sev-4B v0.3.1: policy false alerts 5/140, under the 5 percent ceiling

Browse files
Files changed (8) hide show
  1. README.md +42 -47
  2. adapter_model.safetensors +1 -1
  3. head.pt +2 -2
  4. provenance.json +10 -10
  5. result.json +149 -2726
  6. train.log +11 -85
  7. training_config.json +10 -10
  8. training_metrics.json +7 -7
README.md CHANGED
@@ -19,15 +19,18 @@ tags:
19
  # Sev-4B: security evidence and response policies
20
 
21
  Sev-4B scores supplied answers to questions about security logs and recovered
22
- programs. **v0.3.0-response-policy-research** adds explicit authorization-policy
23
- questions grounded in real SwarmTraces programs. It is a Qwen3.5-4B LoRA adapter
24
- with a decision head. It does not generate text.
25
-
26
- On 175 policy questions from 35 source programs held out of training, accuracy
27
- increases from **127/175, 72.6%, to 147/175, 84.0%** compared with the published
28
- v0.2.0 parent. The user selected this research release with its measured
29
- tradeoffs. **26 of 28 registered checks pass.** The false-alert ceiling and
30
- zero-new-DNS-error check fail. Those results are preserved unchanged.
 
 
 
31
 
32
  ## Use
33
 
@@ -39,7 +42,7 @@ git clone https://github.com/maceip/Sev.git
39
  cd Sev
40
  uv sync --extra serve
41
  uv run python -m kev.serve \
42
- --run macmacmacmac/Sev-4B@v0.3.0-response-policy-research --port 8009
43
  ```
44
 
45
  Send TypeSafe-shaped requests to `POST /v1/systemone`. Supply the observed
@@ -47,8 +50,11 @@ evidence, the applicable policy, and explicit answer choices. The model reads
47
  supplied code without executing it. Its scores support analyst review; they do
48
  not establish that a program ran, identify its human owner, or prove agent origin.
49
 
50
- The checkpoint ships with temperature **1.6817928305074288**. Set
51
- `KEV_TEMPERATURE=1.0` for raw probabilities. Reported evaluation uses
 
 
 
52
  `KEV_BACKEND=torch KEV_DTYPE=fp32 KEV_MERGE=1` on CUDA. Accelerated serving can
53
  differ numerically. The API's `confidence` rescales the largest probability above
54
  chance; it is not an independently verified correctness probability.
@@ -78,8 +84,8 @@ policies and checked labels. Their source text is real recovered code, not a
78
  native execution trace. Source-closure groups separate training, calibration
79
  and development. No locked test was read.
80
 
81
- [Training configuration](training_config.json), [provenance](provenance.json),
82
- and [data lineage](DATA_PROVENANCE.md) pin the inputs. The original
83
  [`v0.1.0-research`](https://huggingface.co/macmacmacmac/Sev-4B/tree/v0.1.0-research)
84
  synthetic-origin preview is a separate historical lineage. The previous
85
  [`v0.2.0-swarmtraces-research`](https://huggingface.co/macmacmacmac/Sev-4B/tree/v0.2.0-swarmtraces-research)
@@ -87,32 +93,23 @@ release remains available unchanged.
87
 
88
  ## Matched development results
89
 
90
- | Policy decision | Published parent | This release |
91
  | --- | ---: | ---: |
92
- | Correct argmax answers | 127/175 | 147/175 |
93
- | Required approvals correct at argmax | 29/35 | 35/35 |
94
- | Permitted answers correct at argmax | 98/140 | 112/140 |
95
- | Required approvals detected at calibrated threshold | 14/35 | 25/35 |
96
- | False alerts at that threshold | 12/140 | 14/140 |
97
- | Calibrated negative log loss | 0.6260 | 0.3743 |
98
- | Calibrated Brier score | 0.4111 | 0.2394 |
99
-
100
- Each model's alert threshold was selected on calibration only, with an empirical
101
- 5% false-alert budget. The parent threshold is 0.887537 and this release's is
102
- 0.844982. Development false alerts increase from 8.6% to **10.0%**, failing the
103
- 5% ceiling. Required-approval recall rises from 40.0% to 71.4%. These are explicit
104
  policy decisions on a development panel, not measured detection rates in live
105
- networks. Always predicting permitted would answer 140/175 correctly and detect
106
- none of the 35 required approvals.
107
 
108
- All eleven extra threshold detections concern `innerText` or `textContent`.
109
- Thirteen of the twenty additional correct argmax answers are console-negative
110
- questions. The panel lacks console-positive and code-execution-positive cases;
111
- it has one onward-request positive. HTML-policy negatives remain weak at 2/21
112
- correct argmax answers, and their threshold false alerts increase from 9/21 to
113
- 12/21. The aggregate does not establish every operation's detection quality.
114
 
115
- | Retained panel | Published parent | This release |
116
  | --- | ---: | ---: |
117
  | Manual SwarmTraces | 65/83 | 66/83 |
118
  | Static SwarmTraces | 345/354 | 347/354 |
@@ -128,12 +125,10 @@ individual answers. GUIDE detector calibrated negative log loss worsens by
128
  0.0116 and Brier by 0.0077. No matched continuation without the new component was
129
  run, so the isolated causal contribution of SwarmTraces is not established.
130
 
131
- The separate DNS diagnostic remains **21/32**, with one repair and one new error.
132
- It contains eight authored summaries and controls inspired by one reported
133
- incident, not raw packet captures or eight independent incidents. It is
134
- evaluation-only. The new error fails the registered zero-regression criterion.
135
- [Policy evaluation](evaluation.json), [DNS evaluation](dns-evaluation.json), and
136
- [registered screen](screen.json) retain the complete comparisons.
137
 
138
  ## Calibration and artifact verification
139
 
@@ -145,20 +140,20 @@ validation. Development and DNS observations were excluded from fitting.
145
 
146
  Only temperature metadata changed when assembling the serving copy. Learned
147
  head tensors, adapter and tokenizer bytes match the evaluated checkpoint.
148
- [Calibration](calibration.json), [integrity evidence](checkpoint-integrity.json),
149
- and [SHA256SUMS](SHA256SUMS) identify the release files.
150
 
151
  Research iteration is paused following this release. The 0.8B and 9B checkpoints
152
  are not updated by this publication.
153
 
154
  ## License and attribution
155
 
156
- Source and adapter/head weights carry Apache-2.0 notices. Preserve [LICENSE](LICENSE),
157
- [NOTICE](NOTICE) and the Qwen [BASE_LICENSE](BASE_LICENSE). Sev builds on Kev by
158
  Jared Palmer and Qwen3.5 by the Qwen team.
159
 
160
  Each dataset retains its own terms. Upstream SwarmTraces reuse terms remain
161
  unverified; the model license does not relicense those artifacts. The collection
162
  includes metadata-only entries, and membership does not imply training use.
163
  The historical synthetic source's notice applies only to that source. See
164
- [data provenance](DATA_PROVENANCE.md).
 
19
  # Sev-4B: security evidence and response policies
20
 
21
  Sev-4B scores supplied answers to questions about security logs and recovered
22
+ programs. **v0.3.1-html-contrast-research** continues
23
+ `v0.3.0-response-policy-research`. It is a Qwen3.5-4B LoRA adapter with a
24
+ decision head. It does not generate text.
25
+
26
+ On 175 policy questions from 35 source programs held out of training, this
27
+ checkpoint answers **158/175, 90.3%**, up from 147/175 on v0.3.0. At the
28
+ calibration-selected alert cut it catches **35/35** required approvals and
29
+ raises **5/140** false alerts, **3.6%**, under the 5% ceiling. v0.3.0 raised
30
+ 14/140 and failed that ceiling. Twelve of those fourteen were allowed HTML
31
+ reads. This pass trains the contrast on the same program: writing `innerText`
32
+ requires approval, and parsing HTML does not. The DNS diagnostic was not rerun.
33
+ The trial's isolation-and-packing check failed. v0.3.0 remains available.
34
 
35
  ## Use
36
 
 
42
  cd Sev
43
  uv sync --extra serve
44
  uv run python -m kev.serve \
45
+ --run macmacmacmac/Sev-4B@v0.3.1-html-contrast-research --port 8009
46
  ```
47
 
48
  Send TypeSafe-shaped requests to `POST /v1/systemone`. Supply the observed
 
50
  supplied code without executing it. Its scores support analyst review; they do
51
  not establish that a program ran, identify its human owner, or prove agent origin.
52
 
53
+ The checkpoint ships with temperature **1.0**. The false-alert cut for the
54
+ policy panel is **1.0** on that raw approval probability: alert only when the
55
+ approval option is certain. That cut was chosen on the 140 policy calibration
56
+ questions, at a 5% false-alert budget, and then applied to development.
57
+ Reported evaluation uses
58
  `KEV_BACKEND=torch KEV_DTYPE=fp32 KEV_MERGE=1` on CUDA. Accelerated serving can
59
  differ numerically. The API's `confidence` rescales the largest probability above
60
  chance; it is not an independently verified correctness probability.
 
84
  native execution trace. Source-closure groups separate training, calibration
85
  and development. No locked test was read.
86
 
87
+ [Training configuration](https://huggingface.co/macmacmacmac/Sev-4B/blob/v0.3.0-response-policy-research/training_config.json), [provenance](https://huggingface.co/macmacmacmac/Sev-4B/blob/v0.3.0-response-policy-research/provenance.json),
88
+ and [data lineage](https://huggingface.co/macmacmacmac/Sev-4B/blob/v0.3.0-response-policy-research/DATA_PROVENANCE.md) pin the inputs. The original
89
  [`v0.1.0-research`](https://huggingface.co/macmacmacmac/Sev-4B/tree/v0.1.0-research)
90
  synthetic-origin preview is a separate historical lineage. The previous
91
  [`v0.2.0-swarmtraces-research`](https://huggingface.co/macmacmacmac/Sev-4B/tree/v0.2.0-swarmtraces-research)
 
93
 
94
  ## Matched development results
95
 
96
+ | Policy decision | v0.3.0 | v0.3.1 |
97
  | --- | ---: | ---: |
98
+ | Correct argmax answers | 147/175 | 158/175 |
99
+ | Required approvals detected at the alert cut | 25/35 | 35/35 |
100
+ | False alerts at that cut | 14/140 | 5/140 |
101
+ | Allowed HTML questions flagged | 12/21 | 0/21 |
102
+
103
+ Each alert cut was selected on calibration only, with a 5% false-alert budget.
104
+ v0.3.1's cut is 1.0 on the raw approval probability. The 5 remaining false
105
+ alerts are 4 console questions and 1 plain-text question. These are explicit
 
 
 
 
106
  policy decisions on a development panel, not measured detection rates in live
107
+ networks.
 
108
 
109
+ The retained-task table and the DNS result below were measured on v0.3.0.
110
+ They were not rerun for v0.3.1.
 
 
 
 
111
 
112
+ | Retained panel | v0.2.0 | v0.3.0, not rerun here |
113
  | --- | ---: | ---: |
114
  | Manual SwarmTraces | 65/83 | 66/83 |
115
  | Static SwarmTraces | 345/354 | 347/354 |
 
125
  0.0116 and Brier by 0.0077. No matched continuation without the new component was
126
  run, so the isolated causal contribution of SwarmTraces is not established.
127
 
128
+ On v0.3.0 the DNS diagnostic was **21/32** and failed the zero-new-error check.
129
+ v0.3.1 did not rerun it.
130
+ [Policy evaluation](https://huggingface.co/macmacmacmac/Sev-4B/blob/v0.3.0-response-policy-research/evaluation.json), [DNS evaluation](https://huggingface.co/macmacmacmac/Sev-4B/blob/v0.3.0-response-policy-research/dns-evaluation.json), and
131
+ [registered screen](https://huggingface.co/macmacmacmac/Sev-4B/blob/v0.3.0-response-policy-research/screen.json) retain the complete comparisons.
 
 
132
 
133
  ## Calibration and artifact verification
134
 
 
140
 
141
  Only temperature metadata changed when assembling the serving copy. Learned
142
  head tensors, adapter and tokenizer bytes match the evaluated checkpoint.
143
+ [Calibration](https://huggingface.co/macmacmacmac/Sev-4B/blob/v0.3.0-response-policy-research/calibration.json), [integrity evidence](https://huggingface.co/macmacmacmac/Sev-4B/blob/v0.3.0-response-policy-research/checkpoint-integrity.json),
144
+ and [SHA256SUMS](https://huggingface.co/macmacmacmac/Sev-4B/blob/v0.3.0-response-policy-research/SHA256SUMS) identify the release files.
145
 
146
  Research iteration is paused following this release. The 0.8B and 9B checkpoints
147
  are not updated by this publication.
148
 
149
  ## License and attribution
150
 
151
+ Source and adapter/head weights carry Apache-2.0 notices. Preserve [LICENSE](https://huggingface.co/macmacmacmac/Sev-4B/blob/v0.3.0-response-policy-research/LICENSE),
152
+ [NOTICE](https://huggingface.co/macmacmacmac/Sev-4B/blob/v0.3.0-response-policy-research/NOTICE) and the Qwen [BASE_LICENSE](https://huggingface.co/macmacmacmac/Sev-4B/blob/v0.3.0-response-policy-research/BASE_LICENSE). Sev builds on Kev by
153
  Jared Palmer and Qwen3.5 by the Qwen team.
154
 
155
  Each dataset retains its own terms. Upstream SwarmTraces reuse terms remain
156
  unverified; the model license does not relicense those artifacts. The collection
157
  includes metadata-only entries, and membership does not imply training use.
158
  The historical synthetic source's notice applies only to that source. See
159
+ [data provenance](https://huggingface.co/macmacmacmac/Sev-4B/blob/v0.3.0-response-policy-research/DATA_PROVENANCE.md).
adapter_model.safetensors CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:fa1916b34fe1eeb10d15c36927c225df43115407d7c97b4f879bba3102c3ed21
3
  size 129924032
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b58b3963175e6c40bd022db5893ebefb6a6f1e1bd1ec11efc6692a089e1dcf51
3
  size 129924032
head.pt CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:12112262e1822fa4ab2981650bd10ceac46929bb29fbbdd843a39ec9de7dbe82
3
- size 5251455
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2724bebcc54d4846194130594b98d3e3b9c632935c0f1d5077e46e23c48b2dc8
3
+ size 5248831
provenance.json CHANGED
@@ -1,8 +1,8 @@
1
  {
2
  "config": {
3
- "epochs": 1,
4
  "seed": 4,
5
- "lr": 2.5e-06,
6
  "lora": 16,
7
  "accum": 2,
8
  "batch": 4,
@@ -23,7 +23,7 @@
23
  "focal_gamma": 0.0,
24
  "base": "Qwen/Qwen3.5-4B-Base",
25
  "base_revision": "1001bb4d826a52d1f399e183466143f4da7b741b",
26
- "init_from": "macmacmacmac/Sev-4B@a1824aefba305fda86e3503b895a7d9b3871f79a",
27
  "lora_targets": "all",
28
  "head_dim": 256,
29
  "weights_dtype": "fp32",
@@ -32,8 +32,8 @@
32
  "option_isolation": 0,
33
  "special_embeddings": 0
34
  },
35
- "config_sha256": "01ce0c9f84977e5c32d40dbd5c012634208c7770615b9e7d9d6f5950c042706f",
36
- "suite_sha256": "3b9a80dfd212b6932b66ea520c35ab1a3b19e132a81af1b6b7b306d776a70a02",
37
  "source_hashes": {
38
  "kev/__init__.py": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
39
  "kev/anchors.py": "6963eafdb276db5a6c94d939eae675c761555553b8d0f963ffd803448beaabbb",
@@ -65,17 +65,17 @@
65
  "pyproject.toml": "39ec40f14c88598dee7ac98fb859a5a44e22337235a3064ebd5e89036ce81e30",
66
  "uv.lock": "a9922dbb89acdef78299fd2b4a8c3f7f0fa1b2bc08b55595b6926fa785a9c466"
67
  },
68
- "git_commit": "740b511bb562cf29430a85457b9e54389a3f21cd",
69
  "platform": "Linux-4.19.0-gvisor-x86_64-with-glibc2.36",
70
  "torch": "2.8.0+cu128",
71
  "device": "cuda",
72
  "gpu": "NVIDIA H100 80GB HBM3",
73
  "legacy_checkpoint": false,
74
  "measured_checkpoint": {
75
- "requested": "/runs/sev-r2-response-policy-4b-v2/00-trial-0/checkpoint",
76
- "resolved": "/runs/sev-r2-response-policy-4b-v2/00-trial-0/checkpoint",
77
- "head_sha256": "506dcd3587847917a288d95757c497c15a2aef8e2d4bc6523b04ad991c961a2b",
78
- "adapter_sha256": "fa1916b34fe1eeb10d15c36927c225df43115407d7c97b4f879bba3102c3ed21",
79
  "inference_temperature": 1.0
80
  }
81
  }
 
1
  {
2
  "config": {
3
+ "epochs": 5,
4
  "seed": 4,
5
+ "lr": 1e-05,
6
  "lora": 16,
7
  "accum": 2,
8
  "batch": 4,
 
23
  "focal_gamma": 0.0,
24
  "base": "Qwen/Qwen3.5-4B-Base",
25
  "base_revision": "1001bb4d826a52d1f399e183466143f4da7b741b",
26
+ "init_from": "macmacmacmac/Sev-4B@v0.3.0-response-policy-research",
27
  "lora_targets": "all",
28
  "head_dim": 256,
29
  "weights_dtype": "fp32",
 
32
  "option_isolation": 0,
33
  "special_embeddings": 0
34
  },
35
+ "config_sha256": "31d8cf5af214ae4ed0012d1e4fea9fd73e207209cfce1f1112107e4aad65cc9f",
36
+ "suite_sha256": "40d3a8b3bd416a3754ecc30e9791a9ebb14a26c8f2a4422057f769a2dbce4dee",
37
  "source_hashes": {
38
  "kev/__init__.py": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
39
  "kev/anchors.py": "6963eafdb276db5a6c94d939eae675c761555553b8d0f963ffd803448beaabbb",
 
65
  "pyproject.toml": "39ec40f14c88598dee7ac98fb859a5a44e22337235a3064ebd5e89036ce81e30",
66
  "uv.lock": "a9922dbb89acdef78299fd2b4a8c3f7f0fa1b2bc08b55595b6926fa785a9c466"
67
  },
68
+ "git_commit": "fcd6756a329f42a515b4b48f83fb8bde9776232f",
69
  "platform": "Linux-4.19.0-gvisor-x86_64-with-glibc2.36",
70
  "torch": "2.8.0+cu128",
71
  "device": "cuda",
72
  "gpu": "NVIDIA H100 80GB HBM3",
73
  "legacy_checkpoint": false,
74
  "measured_checkpoint": {
75
+ "requested": "/runs/sev-html-contrast-4b-v1/00-trial-0/checkpoint",
76
+ "resolved": "/runs/sev-html-contrast-4b-v1/00-trial-0/checkpoint",
77
+ "head_sha256": "2724bebcc54d4846194130594b98d3e3b9c632935c0f1d5077e46e23c48b2dc8",
78
+ "adapter_sha256": "b58b3963175e6c40bd022db5893ebefb6a6f1e1bd1ec11efc6692a089e1dcf51",
79
  "inference_temperature": 1.0
80
  }
81
  }
result.json CHANGED
@@ -1,682 +1,142 @@
1
  {
2
- "objective": -0.17643950559609828,
3
  "paired_flip": null,
4
  "unknowable": null,
5
  "clean": {
6
- "n": 2632,
7
- "nll": 0.4626010690091542,
8
- "acc": 0.7868541033434651,
9
- "ece": 0.0822959829967011,
10
- "brier": 0.2695405506463609,
11
- "mean_conf": 0.8650301490049586,
12
- "confident_error_rate": 0.014437689969604863,
13
- "coverage_at_0_9": 0.6565349544072948,
14
- "accuracy_at_0_9": 0.9780092592592593,
15
- "coverage_at_5pct_error": 0.6808510638297872,
16
- "coverage_at_1pct_error": 0.5828267477203647,
17
- "aurc": 0.04539486327850754,
18
- "error_rate_at_0_9": 0.02199074074074074,
19
- "confidence_bias": 0.0781760456614935,
20
  "top_bins": {
21
  "0.9": {
22
- "n": 1728,
23
- "errors": 38,
24
- "error_rate": 0.02199074074074074
25
  },
26
  "0.95": {
27
- "n": 1675,
28
- "errors": 28,
29
- "error_rate": 0.01671641791044776
30
  },
31
  "0.99": {
32
- "n": 1519,
33
- "errors": 15,
34
- "error_rate": 0.009874917709019092
35
  }
36
  },
37
  "selective": {
38
  "0.5": {
39
- "coverage": 0.5155775075987842,
40
- "accuracy": 0.9918938835666913,
41
- "confidence_cutoff": 0.9953955617749282
42
  },
43
  "0.8": {
44
- "coverage": 0.8028115501519757,
45
- "accuracy": 0.8920965451964032,
46
- "confidence_cutoff": 0.658356020724118
47
  }
48
  }
49
  },
50
  "tasks": {
51
- "excytin_file": {
52
- "n": 80,
53
- "nll": 8.416562133899781e-05,
54
- "acc": 1.0,
55
- "ece": 8.407504508700647e-05,
56
- "brier": 3.608306194877327e-07,
57
- "mean_conf": 0.999915924954913,
58
- "confident_error_rate": 0.0,
59
- "coverage_at_0_9": 1.0,
60
- "accuracy_at_0_9": 1.0,
61
- "coverage_at_5pct_error": 1.0,
62
- "coverage_at_1pct_error": 1.0,
63
- "aurc": 0.0,
64
- "error_rate_at_0_9": 0.0,
65
- "confidence_bias": -8.407504508700647e-05,
66
- "top_bins": {
67
- "0.9": {
68
- "n": 80,
69
- "errors": 0,
70
- "error_rate": 0.0
71
- },
72
- "0.95": {
73
- "n": 80,
74
- "errors": 0,
75
- "error_rate": 0.0
76
- },
77
- "0.99": {
78
- "n": 80,
79
- "errors": 0,
80
- "error_rate": 0.0
81
- }
82
- },
83
- "selective": {
84
- "0.5": {
85
- "coverage": 0.5,
86
- "accuracy": 1.0,
87
- "confidence_cutoff": 0.9999870465451609
88
- },
89
- "0.8": {
90
- "coverage": 0.8,
91
- "accuracy": 1.0,
92
- "confidence_cutoff": 0.9999445146924221
93
- }
94
- }
95
- },
96
- "excytin_parent_join": {
97
- "n": 78,
98
- "nll": 0.0010931532029286512,
99
- "acc": 1.0,
100
- "ece": 0.0010890559164937752,
101
- "brier": 1.524968968159761e-05,
102
- "mean_conf": 0.9989109440835062,
103
- "confident_error_rate": 0.0,
104
- "coverage_at_0_9": 1.0,
105
- "accuracy_at_0_9": 1.0,
106
- "coverage_at_5pct_error": 1.0,
107
- "coverage_at_1pct_error": 1.0,
108
- "aurc": 0.0,
109
- "error_rate_at_0_9": 0.0,
110
- "confidence_bias": -0.0010890559164937752,
111
- "top_bins": {
112
- "0.9": {
113
- "n": 78,
114
- "errors": 0,
115
- "error_rate": 0.0
116
- },
117
- "0.95": {
118
- "n": 78,
119
- "errors": 0,
120
- "error_rate": 0.0
121
- },
122
- "0.99": {
123
- "n": 77,
124
- "errors": 0,
125
- "error_rate": 0.0
126
- }
127
- },
128
- "selective": {
129
- "0.5": {
130
- "coverage": 0.5,
131
- "accuracy": 1.0,
132
- "confidence_cutoff": 0.9997808486535813
133
- },
134
- "0.8": {
135
- "coverage": 0.8076923076923077,
136
- "accuracy": 1.0,
137
- "confidence_cutoff": 0.9985521003494622
138
- }
139
- }
140
- },
141
- "excytin_process": {
142
- "n": 120,
143
- "nll": 0.00038924968205708544,
144
- "acc": 1.0,
145
- "ece": 0.00038892368944332123,
146
- "brier": 1.2813112890567744e-06,
147
- "mean_conf": 0.9996110763105567,
148
- "confident_error_rate": 0.0,
149
- "coverage_at_0_9": 1.0,
150
- "accuracy_at_0_9": 1.0,
151
- "coverage_at_5pct_error": 1.0,
152
- "coverage_at_1pct_error": 1.0,
153
- "aurc": 0.0,
154
- "error_rate_at_0_9": 0.0,
155
- "confidence_bias": -0.00038892368944332123,
156
- "top_bins": {
157
- "0.9": {
158
- "n": 120,
159
- "errors": 0,
160
- "error_rate": 0.0
161
- },
162
- "0.95": {
163
- "n": 120,
164
- "errors": 0,
165
- "error_rate": 0.0
166
- },
167
- "0.99": {
168
- "n": 120,
169
- "errors": 0,
170
- "error_rate": 0.0
171
- }
172
- },
173
- "selective": {
174
- "0.5": {
175
- "coverage": 0.5,
176
- "accuracy": 1.0,
177
- "confidence_cutoff": 0.9999547557532524
178
- },
179
- "0.8": {
180
- "coverage": 0.8,
181
- "accuracy": 1.0,
182
- "confidence_cutoff": 0.9992749850755753
183
- }
184
- }
185
- },
186
- "excytin_signin": {
187
- "n": 120,
188
- "nll": 0.0009466221977245799,
189
- "acc": 1.0,
190
- "ece": 0.0009222842581878554,
191
- "brier": 7.013154556645113e-05,
192
- "mean_conf": 0.9990777157418121,
193
- "confident_error_rate": 0.0,
194
- "coverage_at_0_9": 1.0,
195
- "accuracy_at_0_9": 1.0,
196
- "coverage_at_5pct_error": 1.0,
197
- "coverage_at_1pct_error": 1.0,
198
- "aurc": 0.0,
199
- "error_rate_at_0_9": 0.0,
200
- "confidence_bias": -0.0009222842581878554,
201
- "top_bins": {
202
- "0.9": {
203
- "n": 120,
204
- "errors": 0,
205
- "error_rate": 0.0
206
- },
207
- "0.95": {
208
- "n": 119,
209
- "errors": 0,
210
- "error_rate": 0.0
211
- },
212
- "0.99": {
213
- "n": 119,
214
- "errors": 0,
215
- "error_rate": 0.0
216
- }
217
- },
218
- "selective": {
219
- "0.5": {
220
- "coverage": 0.5,
221
- "accuracy": 1.0,
222
- "confidence_cutoff": 0.9999016374480455
223
- },
224
- "0.8": {
225
- "coverage": 0.8,
226
- "accuracy": 1.0,
227
- "confidence_cutoff": 0.9997216681533668
228
- }
229
- }
230
- },
231
- "sev_guide_detector": {
232
- "n": 1064,
233
- "nll": 0.8605571101399087,
234
- "acc": 0.6005639097744361,
235
- "ece": 0.14734385206430384,
236
- "brier": 0.5008109312564021,
237
- "mean_conf": 0.7382136284198534,
238
- "confident_error_rate": 0.017857142857142856,
239
- "coverage_at_0_9": 0.33270676691729323,
240
- "accuracy_at_0_9": 0.9463276836158192,
241
- "coverage_at_5pct_error": 0.3082706766917293,
242
- "coverage_at_1pct_error": 0.0,
243
- "aurc": 0.19350209200158913,
244
- "error_rate_at_0_9": 0.05367231638418079,
245
- "confidence_bias": 0.1376497186454173,
246
- "top_bins": {
247
- "0.9": {
248
- "n": 354,
249
- "errors": 19,
250
- "error_rate": 0.05367231638418079
251
- },
252
- "0.95": {
253
- "n": 327,
254
- "errors": 13,
255
- "error_rate": 0.039755351681957186
256
- },
257
- "0.99": {
258
- "n": 319,
259
- "errors": 12,
260
- "error_rate": 0.03761755485893417
261
- }
262
- },
263
- "selective": {
264
- "0.5": {
265
- "coverage": 0.5028195488721805,
266
- "accuracy": 0.8205607476635514,
267
- "confidence_cutoff": 0.7059991149306561
268
- },
269
- "0.8": {
270
- "coverage": 0.8035714285714286,
271
- "accuracy": 0.6596491228070176,
272
- "confidence_cutoff": 0.5473316703985754
273
- }
274
- }
275
- },
276
- "sev_guide_triage": {
277
- "n": 231,
278
- "nll": 0.8396428451107271,
279
- "acc": 0.5670995670995671,
280
- "ece": 0.1889188281653689,
281
- "brier": 0.5162927524049745,
282
- "mean_conf": 0.7421494516023849,
283
- "confident_error_rate": 0.0,
284
- "coverage_at_0_9": 0.3593073593073593,
285
- "accuracy_at_0_9": 1.0,
286
- "coverage_at_5pct_error": 0.3722943722943723,
287
- "coverage_at_1pct_error": 0.36363636363636365,
288
- "aurc": 0.20987091476964936,
289
- "error_rate_at_0_9": 0.0,
290
- "confidence_bias": 0.17504988450281778,
291
- "top_bins": {
292
- "0.9": {
293
- "n": 83,
294
- "errors": 0,
295
- "error_rate": 0.0
296
- },
297
- "0.95": {
298
- "n": 80,
299
- "errors": 0,
300
- "error_rate": 0.0
301
- },
302
- "0.99": {
303
- "n": 14,
304
- "errors": 0,
305
- "error_rate": 0.0
306
- }
307
- },
308
- "selective": {
309
- "0.5": {
310
- "coverage": 0.5627705627705628,
311
- "accuracy": 0.7153846153846154,
312
- "confidence_cutoff": 0.7564145922660828
313
- },
314
- "0.8": {
315
- "coverage": 0.8051948051948052,
316
- "accuracy": 0.5806451612903226,
317
- "confidence_cutoff": 0.51023405030021
318
- }
319
- }
320
- },
321
  "sev_swarmtraces_response_policy": {
322
  "n": 175,
323
- "nll": 0.49844721765472333,
324
- "acc": 0.84,
325
- "ece": 0.1173563845984423,
326
- "brier": 0.26967228692055817,
327
- "mean_conf": 0.955988843320762,
328
- "confident_error_rate": 0.10285714285714286,
329
- "coverage_at_0_9": 0.8742857142857143,
330
- "accuracy_at_0_9": 0.8823529411764706,
331
- "coverage_at_5pct_error": 0.6114285714285714,
332
- "coverage_at_1pct_error": 0.44,
333
- "aurc": 0.044632779737602504,
334
- "error_rate_at_0_9": 0.11764705882352941,
335
- "confidence_bias": 0.11598884332076198,
336
  "top_bins": {
337
  "0.9": {
338
- "n": 153,
339
- "errors": 18,
340
- "error_rate": 0.11764705882352941
341
  },
342
  "0.95": {
343
- "n": 140,
344
  "errors": 14,
345
- "error_rate": 0.1
346
  },
347
  "0.99": {
348
- "n": 92,
349
- "errors": 3,
350
- "error_rate": 0.03260869565217391
351
  }
352
  },
353
  "selective": {
354
  "0.5": {
355
  "coverage": 0.5028571428571429,
356
- "accuracy": 0.9772727272727273,
357
- "confidence_cutoff": 0.9904485381783863
358
  },
359
  "0.8": {
360
  "coverage": 0.8,
361
- "accuracy": 0.9,
362
- "confidence_cutoff": 0.9516673045977523
363
- }
364
- }
365
- },
366
- "sev_swarmtraces_static": {
367
- "n": 354,
368
- "nll": 0.04944532768977841,
369
- "acc": 0.980225988700565,
370
- "ece": 0.01487544035693287,
371
- "brier": 0.024608535342050798,
372
- "mean_conf": 0.9802643779708107,
373
- "confident_error_rate": 0.002824858757062147,
374
- "coverage_at_0_9": 0.9463276836158192,
375
- "accuracy_at_0_9": 0.9970149253731343,
376
- "coverage_at_5pct_error": 1.0,
377
- "coverage_at_1pct_error": 0.9858757062146892,
378
- "aurc": 0.0005032049402250012,
379
- "error_rate_at_0_9": 0.0029850746268656717,
380
- "confidence_bias": 3.8389270245686724e-05,
381
- "top_bins": {
382
- "0.9": {
383
- "n": 335,
384
- "errors": 1,
385
- "error_rate": 0.0029850746268656717
386
- },
387
- "0.95": {
388
- "n": 330,
389
- "errors": 1,
390
- "error_rate": 0.0030303030303030303
391
- },
392
- "0.99": {
393
- "n": 308,
394
- "errors": 0,
395
- "error_rate": 0.0
396
- }
397
- },
398
- "selective": {
399
- "0.5": {
400
- "coverage": 0.5,
401
- "accuracy": 1.0,
402
- "confidence_cutoff": 0.9999276209511099
403
- },
404
- "0.8": {
405
- "coverage": 0.8022598870056498,
406
- "accuracy": 1.0,
407
- "confidence_cutoff": 0.997796467038603
408
- }
409
- }
410
- },
411
- "sev_sysmon_native_file": {
412
- "n": 62,
413
- "nll": 0.008428780795422771,
414
- "acc": 1.0,
415
- "ece": 0.007810434776248104,
416
- "brier": 0.0017443201233063612,
417
- "mean_conf": 0.9921895652237519,
418
- "confident_error_rate": 0.0,
419
- "coverage_at_0_9": 0.9838709677419355,
420
- "accuracy_at_0_9": 1.0,
421
- "coverage_at_5pct_error": 1.0,
422
- "coverage_at_1pct_error": 1.0,
423
- "aurc": 0.0,
424
- "error_rate_at_0_9": 0.0,
425
- "confidence_bias": -0.0078104347762480986,
426
- "top_bins": {
427
- "0.9": {
428
- "n": 61,
429
- "errors": 0,
430
- "error_rate": 0.0
431
- },
432
- "0.95": {
433
- "n": 59,
434
- "errors": 0,
435
- "error_rate": 0.0
436
- },
437
- "0.99": {
438
- "n": 57,
439
- "errors": 0,
440
- "error_rate": 0.0
441
- }
442
- },
443
- "selective": {
444
- "0.5": {
445
- "coverage": 0.5,
446
- "accuracy": 1.0,
447
- "confidence_cutoff": 0.9991686260052061
448
- },
449
- "0.8": {
450
- "coverage": 0.8064516129032258,
451
- "accuracy": 1.0,
452
- "confidence_cutoff": 0.9962482122843107
453
- }
454
- }
455
- },
456
- "sev_sysmon_native_join": {
457
- "n": 120,
458
- "nll": 0.00039767294618876087,
459
- "acc": 1.0,
460
- "ece": 0.0003974740966455892,
461
- "brier": 7.310901229709924e-07,
462
- "mean_conf": 0.9996025259033544,
463
- "confident_error_rate": 0.0,
464
- "coverage_at_0_9": 1.0,
465
- "accuracy_at_0_9": 1.0,
466
- "coverage_at_5pct_error": 1.0,
467
- "coverage_at_1pct_error": 1.0,
468
- "aurc": 0.0,
469
- "error_rate_at_0_9": 0.0,
470
- "confidence_bias": -0.0003974740966455892,
471
- "top_bins": {
472
- "0.9": {
473
- "n": 120,
474
- "errors": 0,
475
- "error_rate": 0.0
476
- },
477
- "0.95": {
478
- "n": 120,
479
- "errors": 0,
480
- "error_rate": 0.0
481
- },
482
- "0.99": {
483
- "n": 120,
484
- "errors": 0,
485
- "error_rate": 0.0
486
- }
487
- },
488
- "selective": {
489
- "0.5": {
490
- "coverage": 0.5,
491
- "accuracy": 1.0,
492
- "confidence_cutoff": 0.9997660582574378
493
- },
494
- "0.8": {
495
- "coverage": 0.8,
496
- "accuracy": 1.0,
497
- "confidence_cutoff": 0.9994606035447323
498
- }
499
- }
500
- },
501
- "sev_sysmon_native_other_effect": {
502
- "n": 88,
503
- "nll": 0.005466714567741097,
504
- "acc": 1.0,
505
- "ece": 0.005048372154544871,
506
- "brier": 0.0014126265728775101,
507
- "mean_conf": 0.994951627845455,
508
- "confident_error_rate": 0.0,
509
- "coverage_at_0_9": 0.9886363636363636,
510
- "accuracy_at_0_9": 1.0,
511
- "coverage_at_5pct_error": 1.0,
512
- "coverage_at_1pct_error": 1.0,
513
- "aurc": 0.0,
514
- "error_rate_at_0_9": 0.0,
515
- "confidence_bias": -0.005048372154545011,
516
- "top_bins": {
517
- "0.9": {
518
- "n": 87,
519
- "errors": 0,
520
- "error_rate": 0.0
521
- },
522
- "0.95": {
523
- "n": 86,
524
- "errors": 0,
525
- "error_rate": 0.0
526
- },
527
- "0.99": {
528
- "n": 83,
529
- "errors": 0,
530
- "error_rate": 0.0
531
- }
532
- },
533
- "selective": {
534
- "0.5": {
535
- "coverage": 0.5,
536
- "accuracy": 1.0,
537
- "confidence_cutoff": 0.9997677646417501
538
- },
539
- "0.8": {
540
- "coverage": 0.8068181818181818,
541
- "accuracy": 1.0,
542
- "confidence_cutoff": 0.9978176224669177
543
- }
544
- }
545
- },
546
- "sev_sysmon_native_process": {
547
- "n": 74,
548
- "nll": 0.004798899209615666,
549
- "acc": 1.0,
550
- "ece": 0.004644321057477101,
551
- "brier": 0.0005716684001551928,
552
- "mean_conf": 0.995355678942523,
553
- "confident_error_rate": 0.0,
554
- "coverage_at_0_9": 0.9864864864864865,
555
- "accuracy_at_0_9": 1.0,
556
- "coverage_at_5pct_error": 1.0,
557
- "coverage_at_1pct_error": 1.0,
558
- "aurc": 0.0,
559
- "error_rate_at_0_9": 0.0,
560
- "confidence_bias": -0.00464432105747703,
561
- "top_bins": {
562
- "0.9": {
563
- "n": 73,
564
- "errors": 0,
565
- "error_rate": 0.0
566
- },
567
- "0.95": {
568
- "n": 72,
569
- "errors": 0,
570
- "error_rate": 0.0
571
- },
572
- "0.99": {
573
- "n": 68,
574
- "errors": 0,
575
- "error_rate": 0.0
576
- }
577
- },
578
- "selective": {
579
- "0.5": {
580
- "coverage": 0.5,
581
- "accuracy": 1.0,
582
- "confidence_cutoff": 0.9999268900168323
583
- },
584
- "0.8": {
585
- "coverage": 0.8108108108108109,
586
- "accuracy": 1.0,
587
- "confidence_cutoff": 0.9973453419434639
588
- }
589
- }
590
- },
591
- "sev_sysmon_native_registry": {
592
- "n": 66,
593
- "nll": 0.024015813931122083,
594
- "acc": 0.9848484848484849,
595
- "ece": 0.016913465896023366,
596
- "brier": 0.01690246747159733,
597
- "mean_conf": 0.9866786138448916,
598
- "confident_error_rate": 0.0,
599
- "coverage_at_0_9": 0.9696969696969697,
600
- "accuracy_at_0_9": 1.0,
601
- "coverage_at_5pct_error": 1.0,
602
- "coverage_at_1pct_error": 0.9696969696969697,
603
- "aurc": 0.00046266864448682633,
604
- "error_rate_at_0_9": 0.0,
605
- "confidence_bias": 0.0018301289964067768,
606
- "top_bins": {
607
- "0.9": {
608
- "n": 64,
609
- "errors": 0,
610
- "error_rate": 0.0
611
- },
612
- "0.95": {
613
- "n": 64,
614
- "errors": 0,
615
- "error_rate": 0.0
616
- },
617
- "0.99": {
618
- "n": 62,
619
- "errors": 0,
620
- "error_rate": 0.0
621
- }
622
- },
623
- "selective": {
624
- "0.5": {
625
- "coverage": 0.5,
626
- "accuracy": 1.0,
627
- "confidence_cutoff": 0.9998656142719909
628
- },
629
- "0.8": {
630
- "coverage": 0.803030303030303,
631
- "accuracy": 1.0,
632
- "confidence_cutoff": 0.9982153954384291
633
  }
634
  }
635
  }
636
  },
637
  "variants": {
638
  "clean": {
639
- "n": 2632,
640
- "nll": 0.4626010690091542,
641
- "acc": 0.7868541033434651,
642
- "ece": 0.0822959829967011,
643
- "brier": 0.2695405506463609,
644
- "mean_conf": 0.8650301490049586,
645
- "confident_error_rate": 0.014437689969604863,
646
- "coverage_at_0_9": 0.6565349544072948,
647
- "accuracy_at_0_9": 0.9780092592592593,
648
- "coverage_at_5pct_error": 0.6808510638297872,
649
- "coverage_at_1pct_error": 0.5828267477203647,
650
- "aurc": 0.04539486327850754,
651
- "error_rate_at_0_9": 0.02199074074074074,
652
- "confidence_bias": 0.0781760456614935,
653
  "top_bins": {
654
  "0.9": {
655
- "n": 1728,
656
- "errors": 38,
657
- "error_rate": 0.02199074074074074
658
  },
659
  "0.95": {
660
- "n": 1675,
661
- "errors": 28,
662
- "error_rate": 0.01671641791044776
663
  },
664
  "0.99": {
665
- "n": 1519,
666
- "errors": 15,
667
- "error_rate": 0.009874917709019092
668
  }
669
  },
670
  "selective": {
671
  "0.5": {
672
- "coverage": 0.5155775075987842,
673
- "accuracy": 0.9918938835666913,
674
- "confidence_cutoff": 0.9953955617749282
675
  },
676
  "0.8": {
677
- "coverage": 0.8028115501519757,
678
- "accuracy": 0.8920965451964032,
679
- "confidence_cutoff": 0.658356020724118
680
  }
681
  }
682
  }
@@ -687,49 +147,49 @@
687
  "mean_max_delta": null,
688
  "flip_rate": null
689
  },
690
- "temperature": 1.6817928305074288,
691
  "calibrated_clean": {
692
- "n": 2632,
693
- "nll": 0.42892992382965894,
694
- "acc": 0.7868541033434651,
695
- "ece": 0.04404789769068498,
696
- "brier": 0.25086130169182097,
697
- "mean_conf": 0.8094973007371603,
698
- "confident_error_rate": 0.007978723404255319,
699
- "coverage_at_0_9": 0.6196808510638298,
700
- "accuracy_at_0_9": 0.9871244635193133,
701
- "coverage_at_5pct_error": 0.6918693009118541,
702
- "coverage_at_1pct_error": 0.5395136778115501,
703
- "aurc": 0.045402363700726256,
704
- "error_rate_at_0_9": 0.012875536480686695,
705
- "confidence_bias": 0.022643197393695247,
706
  "top_bins": {
707
  "0.9": {
708
- "n": 1631,
709
- "errors": 21,
710
- "error_rate": 0.012875536480686695
711
  },
712
  "0.95": {
713
- "n": 1153,
714
- "errors": 0,
715
- "error_rate": 0.0
716
  },
717
  "0.99": {
718
- "n": 776,
719
  "errors": 0,
720
- "error_rate": 0.0
721
  }
722
  },
723
  "selective": {
724
  "0.5": {
725
- "coverage": 0.5208966565349544,
726
- "accuracy": 0.9912472647702407,
727
- "confidence_cutoff": 0.9486799539412789
728
  },
729
  "0.8": {
730
- "coverage": 0.8111702127659575,
731
- "accuracy": 0.8861826697892272,
732
- "confidence_cutoff": 0.52766162421842
733
  }
734
  }
735
  },
@@ -747,2062 +207,28 @@
747
  "returned_zeros": 0
748
  },
749
  "coverage": {
750
- "requested_records": 1828,
751
- "requested_questions": 2632,
752
- "evaluated_records": 1828,
753
- "evaluated_questions": 2632,
754
  "rejected_records": 0,
755
  "truncated_records": 0
756
  },
757
  "latency_ms": {
758
- "median": 54.25747899994349,
759
- "p95": 161.14951469993457
760
  },
761
  "calibration": {
762
  "inference_temperature": 1.0,
763
- "additional_temperature": 1.6817928305074288,
764
  "logits_recorded": true
765
  },
766
- "transfer": {
767
- "objective": -0.4988755241804635,
768
- "paired_flip": {
769
- "pairs": 16,
770
- "flip_rate": 0.9375,
771
- "both_correct_rate": 0.9375,
772
- "invariant_pairs": 8,
773
- "invariance_rate": 1.0,
774
- "invariant_both_correct_rate": 0.875
775
- },
776
- "unknowable": null,
777
- "clean": {
778
- "n": 243,
779
- "nll": 0.6065943824685833,
780
- "acc": 0.8683127572016461,
781
- "ece": 0.09195653550207036,
782
- "brier": 0.23499321102178952,
783
- "mean_conf": 0.9525644631029061,
784
- "confident_error_rate": 0.09053497942386832,
785
- "coverage_at_0_9": 0.8600823045267489,
786
- "accuracy_at_0_9": 0.8947368421052632,
787
- "coverage_at_5pct_error": 0.6872427983539094,
788
- "coverage_at_1pct_error": 0.42386831275720166,
789
- "aurc": 0.034793809228677,
790
- "error_rate_at_0_9": 0.10526315789473684,
791
- "confidence_bias": 0.08425170590125997,
792
- "top_bins": {
793
- "0.9": {
794
- "n": 209,
795
- "errors": 22,
796
- "error_rate": 0.10526315789473684
797
- },
798
- "0.95": {
799
- "n": 193,
800
- "errors": 13,
801
- "error_rate": 0.06735751295336788
802
- },
803
- "0.99": {
804
- "n": 154,
805
- "errors": 4,
806
- "error_rate": 0.025974025974025976
807
- }
808
- },
809
- "selective": {
810
- "0.5": {
811
- "coverage": 0.5020576131687243,
812
- "accuracy": 0.9836065573770492,
813
- "confidence_cutoff": 0.9985267483650163
814
- },
815
- "0.8": {
816
- "coverage": 0.8024691358024691,
817
- "accuracy": 0.9230769230769231,
818
- "confidence_cutoff": 0.946113674977154
819
- }
820
- }
821
- },
822
- "tasks": {
823
- "composition_held_and_or": {
824
- "n": 32,
825
- "nll": 0.24171629359001318,
826
- "acc": 0.90625,
827
- "ece": 0.08267706695037225,
828
- "brier": 0.1580607438572586,
829
- "mean_conf": 0.9889270669503722,
830
- "confident_error_rate": 0.09375,
831
- "coverage_at_0_9": 1.0,
832
- "accuracy_at_0_9": 0.90625,
833
- "coverage_at_5pct_error": 0.9375,
834
- "coverage_at_1pct_error": 0.90625,
835
- "aurc": 0.0069010416666666664,
836
- "error_rate_at_0_9": 0.09375,
837
- "confidence_bias": 0.08267706695037225,
838
- "top_bins": {
839
- "0.9": {
840
- "n": 32,
841
- "errors": 3,
842
- "error_rate": 0.09375
843
- },
844
- "0.95": {
845
- "n": 30,
846
- "errors": 1,
847
- "error_rate": 0.03333333333333333
848
- },
849
- "0.99": {
850
- "n": 26,
851
- "errors": 0,
852
- "error_rate": 0.0
853
- }
854
- },
855
- "selective": {
856
- "0.5": {
857
- "coverage": 0.5,
858
- "accuracy": 1.0,
859
- "confidence_cutoff": 0.9999364856833449
860
- },
861
- "0.8": {
862
- "coverage": 0.8125,
863
- "accuracy": 1.0,
864
- "confidence_cutoff": 0.9996934954887139
865
- }
866
- }
867
- },
868
- "contrastive_authorization": {
869
- "n": 16,
870
- "nll": 1.502653141036028e-05,
871
- "acc": 1.0,
872
- "ece": 1.50264042420023e-05,
873
- "brier": 5.139828295030656e-10,
874
- "mean_conf": 0.999984973595758,
875
- "confident_error_rate": 0.0,
876
- "coverage_at_0_9": 1.0,
877
- "accuracy_at_0_9": 1.0,
878
- "coverage_at_5pct_error": 1.0,
879
- "coverage_at_1pct_error": 1.0,
880
- "aurc": 0.0,
881
- "error_rate_at_0_9": 0.0,
882
- "confidence_bias": -1.50264042420023e-05,
883
- "top_bins": {
884
- "0.9": {
885
- "n": 16,
886
- "errors": 0,
887
- "error_rate": 0.0
888
- },
889
- "0.95": {
890
- "n": 16,
891
- "errors": 0,
892
- "error_rate": 0.0
893
- },
894
- "0.99": {
895
- "n": 16,
896
- "errors": 0,
897
- "error_rate": 0.0
898
- }
899
- },
900
- "selective": {
901
- "0.5": {
902
- "coverage": 0.5,
903
- "accuracy": 1.0,
904
- "confidence_cutoff": 0.9999862890472545
905
- },
906
- "0.8": {
907
- "coverage": 0.8125,
908
- "accuracy": 1.0,
909
- "confidence_cutoff": 0.9999813198967473
910
- }
911
- }
912
- },
913
- "emotion": {
914
- "n": 8,
915
- "nll": 1.4194273414243654,
916
- "acc": 0.625,
917
- "ece": 0.3420785503418361,
918
- "brier": 0.6845960520300963,
919
- "mean_conf": 0.967078550341836,
920
- "confident_error_rate": 0.25,
921
- "coverage_at_0_9": 0.875,
922
- "accuracy_at_0_9": 0.7142857142857143,
923
- "coverage_at_5pct_error": 0.5,
924
- "coverage_at_1pct_error": 0.5,
925
- "aurc": 0.12842261904761904,
926
- "error_rate_at_0_9": 0.2857142857142857,
927
- "confidence_bias": 0.342078550341836,
928
- "top_bins": {
929
- "0.9": {
930
- "n": 7,
931
- "errors": 2,
932
- "error_rate": 0.2857142857142857
933
- },
934
- "0.95": {
935
- "n": 5,
936
- "errors": 1,
937
- "error_rate": 0.2
938
- },
939
- "0.99": {
940
- "n": 3,
941
- "errors": 0,
942
- "error_rate": 0.0
943
- }
944
- },
945
- "selective": {
946
- "0.5": {
947
- "coverage": 0.5,
948
- "accuracy": 1.0,
949
- "confidence_cutoff": 0.9869563807629563
950
- },
951
- "0.8": {
952
- "coverage": 0.875,
953
- "accuracy": 0.7142857142857143,
954
- "confidence_cutoff": 0.936063568486461
955
- }
956
- }
957
- },
958
- "mmlu": {
959
- "n": 8,
960
- "nll": 1.3129749769552814,
961
- "acc": 0.625,
962
- "ece": 0.15765322690405312,
963
- "brier": 0.6484351711065982,
964
- "mean_conf": 0.7567759888850003,
965
- "confident_error_rate": 0.125,
966
- "coverage_at_0_9": 0.5,
967
- "accuracy_at_0_9": 0.75,
968
- "coverage_at_5pct_error": 0.375,
969
- "coverage_at_1pct_error": 0.375,
970
- "aurc": 0.22336309523809525,
971
- "error_rate_at_0_9": 0.25,
972
- "confidence_bias": 0.1317759888850003,
973
- "top_bins": {
974
- "0.9": {
975
- "n": 4,
976
- "errors": 1,
977
- "error_rate": 0.25
978
- },
979
- "0.95": {
980
- "n": 3,
981
- "errors": 0,
982
- "error_rate": 0.0
983
- },
984
- "0.99": {
985
- "n": 1,
986
- "errors": 0,
987
- "error_rate": 0.0
988
- }
989
- },
990
- "selective": {
991
- "0.5": {
992
- "coverage": 0.5,
993
- "accuracy": 0.75,
994
- "confidence_cutoff": 0.9327223999584052
995
- },
996
- "0.8": {
997
- "coverage": 0.875,
998
- "accuracy": 0.5714285714285714,
999
- "confidence_cutoff": 0.485476400811401
1000
- }
1001
- }
1002
- },
1003
- "otrf_sysmon_file": {
1004
- "n": 16,
1005
- "nll": 0.008605175298168915,
1006
- "acc": 1.0,
1007
- "ece": 0.008437849080092708,
1008
- "brier": 0.0005211762382657875,
1009
- "mean_conf": 0.9915621509199073,
1010
- "confident_error_rate": 0.0,
1011
- "coverage_at_0_9": 1.0,
1012
- "accuracy_at_0_9": 1.0,
1013
- "coverage_at_5pct_error": 1.0,
1014
- "coverage_at_1pct_error": 1.0,
1015
- "aurc": 0.0,
1016
- "error_rate_at_0_9": 0.0,
1017
- "confidence_bias": -0.008437849080092708,
1018
- "top_bins": {
1019
- "0.9": {
1020
- "n": 16,
1021
- "errors": 0,
1022
- "error_rate": 0.0
1023
- },
1024
- "0.95": {
1025
- "n": 15,
1026
- "errors": 0,
1027
- "error_rate": 0.0
1028
- },
1029
- "0.99": {
1030
- "n": 13,
1031
- "errors": 0,
1032
- "error_rate": 0.0
1033
- }
1034
- },
1035
- "selective": {
1036
- "0.5": {
1037
- "coverage": 0.5,
1038
- "accuracy": 1.0,
1039
- "confidence_cutoff": 0.9989837483137961
1040
- },
1041
- "0.8": {
1042
- "coverage": 0.8125,
1043
- "accuracy": 1.0,
1044
- "confidence_cutoff": 0.9925042879134804
1045
- }
1046
- }
1047
- },
1048
- "otrf_sysmon_join": {
1049
- "n": 16,
1050
- "nll": 0.12967885980971613,
1051
- "acc": 0.9375,
1052
- "ece": 0.0857233928404766,
1053
- "brier": 0.08146985002971384,
1054
- "mean_conf": 0.946506108546294,
1055
- "confident_error_rate": 0.0,
1056
- "coverage_at_0_9": 0.6875,
1057
- "accuracy_at_0_9": 1.0,
1058
- "coverage_at_5pct_error": 0.9375,
1059
- "coverage_at_1pct_error": 0.9375,
1060
- "aurc": 0.00390625,
1061
- "error_rate_at_0_9": 0.0,
1062
- "confidence_bias": 0.009006108546294023,
1063
- "top_bins": {
1064
- "0.9": {
1065
- "n": 11,
1066
- "errors": 0,
1067
- "error_rate": 0.0
1068
- },
1069
- "0.95": {
1070
- "n": 11,
1071
- "errors": 0,
1072
- "error_rate": 0.0
1073
- },
1074
- "0.99": {
1075
- "n": 9,
1076
- "errors": 0,
1077
- "error_rate": 0.0
1078
- }
1079
- },
1080
- "selective": {
1081
- "0.5": {
1082
- "coverage": 0.5,
1083
- "accuracy": 1.0,
1084
- "confidence_cutoff": 0.9970209747654696
1085
- },
1086
- "0.8": {
1087
- "coverage": 0.8125,
1088
- "accuracy": 1.0,
1089
- "confidence_cutoff": 0.8702734279589676
1090
- }
1091
- }
1092
- },
1093
- "otrf_sysmon_process": {
1094
- "n": 24,
1095
- "nll": 0.16156109121321785,
1096
- "acc": 0.9583333333333334,
1097
- "ece": 0.05273169620250518,
1098
- "brier": 0.0899182770612567,
1099
- "mean_conf": 0.9768730534919899,
1100
- "confident_error_rate": 0.041666666666666664,
1101
- "coverage_at_0_9": 0.9583333333333334,
1102
- "accuracy_at_0_9": 0.9565217391304348,
1103
- "coverage_at_5pct_error": 1.0,
1104
- "coverage_at_1pct_error": 0.9166666666666666,
1105
- "aurc": 0.0035477053140096615,
1106
- "error_rate_at_0_9": 0.043478260869565216,
1107
- "confidence_bias": 0.01853972015865657,
1108
- "top_bins": {
1109
- "0.9": {
1110
- "n": 23,
1111
- "errors": 1,
1112
- "error_rate": 0.043478260869565216
1113
- },
1114
- "0.95": {
1115
- "n": 22,
1116
- "errors": 0,
1117
- "error_rate": 0.0
1118
- },
1119
- "0.99": {
1120
- "n": 19,
1121
- "errors": 0,
1122
- "error_rate": 0.0
1123
- }
1124
- },
1125
- "selective": {
1126
- "0.5": {
1127
- "coverage": 0.5,
1128
- "accuracy": 1.0,
1129
- "confidence_cutoff": 0.9994396085787187
1130
- },
1131
- "0.8": {
1132
- "coverage": 0.8333333333333334,
1133
- "accuracy": 1.0,
1134
- "confidence_cutoff": 0.9832554759400715
1135
- }
1136
- }
1137
- },
1138
- "otrf_sysmon_registry": {
1139
- "n": 8,
1140
- "nll": 0.010723606364118954,
1141
- "acc": 1.0,
1142
- "ece": 0.010542385937563847,
1143
- "brier": 0.0006959554565011166,
1144
- "mean_conf": 0.9894576140624362,
1145
- "confident_error_rate": 0.0,
1146
- "coverage_at_0_9": 1.0,
1147
- "accuracy_at_0_9": 1.0,
1148
- "coverage_at_5pct_error": 1.0,
1149
- "coverage_at_1pct_error": 1.0,
1150
- "aurc": 0.0,
1151
- "error_rate_at_0_9": 0.0,
1152
- "confidence_bias": -0.010542385937563847,
1153
- "top_bins": {
1154
- "0.9": {
1155
- "n": 8,
1156
- "errors": 0,
1157
- "error_rate": 0.0
1158
- },
1159
- "0.95": {
1160
- "n": 8,
1161
- "errors": 0,
1162
- "error_rate": 0.0
1163
- },
1164
- "0.99": {
1165
- "n": 5,
1166
- "errors": 0,
1167
- "error_rate": 0.0
1168
- }
1169
- },
1170
- "selective": {
1171
- "0.5": {
1172
- "coverage": 0.5,
1173
- "accuracy": 1.0,
1174
- "confidence_cutoff": 0.9972651302471073
1175
- },
1176
- "0.8": {
1177
- "coverage": 0.875,
1178
- "accuracy": 1.0,
1179
- "confidence_cutoff": 0.9840692701748682
1180
- }
1181
- }
1182
- },
1183
- "paws": {
1184
- "n": 8,
1185
- "nll": 1.0368545966100273,
1186
- "acc": 0.75,
1187
- "ece": 0.22498828525853157,
1188
- "brier": 0.34163600039007924,
1189
- "mean_conf": 0.9122532529823204,
1190
- "confident_error_rate": 0.125,
1191
- "coverage_at_0_9": 0.75,
1192
- "accuracy_at_0_9": 0.8333333333333334,
1193
- "coverage_at_5pct_error": 0.5,
1194
- "coverage_at_1pct_error": 0.5,
1195
- "aurc": 0.09494047619047619,
1196
- "error_rate_at_0_9": 0.16666666666666666,
1197
- "confidence_bias": 0.1622532529823204,
1198
- "top_bins": {
1199
- "0.9": {
1200
- "n": 6,
1201
- "errors": 1,
1202
- "error_rate": 0.16666666666666666
1203
- },
1204
- "0.95": {
1205
- "n": 6,
1206
- "errors": 1,
1207
- "error_rate": 0.16666666666666666
1208
- },
1209
- "0.99": {
1210
- "n": 6,
1211
- "errors": 1,
1212
- "error_rate": 0.16666666666666666
1213
- }
1214
- },
1215
- "selective": {
1216
- "0.5": {
1217
- "coverage": 0.5,
1218
- "accuracy": 1.0,
1219
- "confidence_cutoff": 0.9992691391151949
1220
- },
1221
- "0.8": {
1222
- "coverage": 0.875,
1223
- "accuracy": 0.8571428571428571,
1224
- "confidence_cutoff": 0.749059870895155
1225
- }
1226
- }
1227
- },
1228
- "qnli": {
1229
- "n": 8,
1230
- "nll": 0.005181466224742151,
1231
- "acc": 1.0,
1232
- "ece": 0.005129334884136472,
1233
- "brier": 0.00020524982613720548,
1234
- "mean_conf": 0.9948706651158635,
1235
- "confident_error_rate": 0.0,
1236
- "coverage_at_0_9": 1.0,
1237
- "accuracy_at_0_9": 1.0,
1238
- "coverage_at_5pct_error": 1.0,
1239
- "coverage_at_1pct_error": 1.0,
1240
- "aurc": 0.0,
1241
- "error_rate_at_0_9": 0.0,
1242
- "confidence_bias": -0.005129334884136472,
1243
- "top_bins": {
1244
- "0.9": {
1245
- "n": 8,
1246
- "errors": 0,
1247
- "error_rate": 0.0
1248
- },
1249
- "0.95": {
1250
- "n": 8,
1251
- "errors": 0,
1252
- "error_rate": 0.0
1253
- },
1254
- "0.99": {
1255
- "n": 6,
1256
- "errors": 0,
1257
- "error_rate": 0.0
1258
- }
1259
- },
1260
- "selective": {
1261
- "0.5": {
1262
- "coverage": 0.5,
1263
- "accuracy": 1.0,
1264
- "confidence_cutoff": 0.999952371295075
1265
- },
1266
- "0.8": {
1267
- "coverage": 0.875,
1268
- "accuracy": 1.0,
1269
- "confidence_cutoff": 0.98872862957213
1270
- }
1271
- }
1272
- },
1273
- "sciq": {
1274
- "n": 8,
1275
- "nll": 0.00023928943016782004,
1276
- "acc": 1.0,
1277
- "ece": 0.0002391650142747892,
1278
- "brier": 4.955209174865095e-07,
1279
- "mean_conf": 0.9997608349857252,
1280
- "confident_error_rate": 0.0,
1281
- "coverage_at_0_9": 1.0,
1282
- "accuracy_at_0_9": 1.0,
1283
- "coverage_at_5pct_error": 1.0,
1284
- "coverage_at_1pct_error": 1.0,
1285
- "aurc": 0.0,
1286
- "error_rate_at_0_9": 0.0,
1287
- "confidence_bias": -0.0002391650142747892,
1288
- "top_bins": {
1289
- "0.9": {
1290
- "n": 8,
1291
- "errors": 0,
1292
- "error_rate": 0.0
1293
- },
1294
- "0.95": {
1295
- "n": 8,
1296
- "errors": 0,
1297
- "error_rate": 0.0
1298
- },
1299
- "0.99": {
1300
- "n": 8,
1301
- "errors": 0,
1302
- "error_rate": 0.0
1303
- }
1304
- },
1305
- "selective": {
1306
- "0.5": {
1307
- "coverage": 0.5,
1308
- "accuracy": 1.0,
1309
- "confidence_cutoff": 0.9999911919978735
1310
- },
1311
- "0.8": {
1312
- "coverage": 0.875,
1313
- "accuracy": 1.0,
1314
- "confidence_cutoff": 0.9994259298933852
1315
- }
1316
- }
1317
- },
1318
- "sev_swarmtraces_manual_credential": {
1319
- "n": 40,
1320
- "nll": 0.9530052125004369,
1321
- "acc": 0.825,
1322
- "ece": 0.17316335764162732,
1323
- "brier": 0.3580423587545788,
1324
- "mean_conf": 0.9202306277592299,
1325
- "confident_error_rate": 0.15,
1326
- "coverage_at_0_9": 0.75,
1327
- "accuracy_at_0_9": 0.8,
1328
- "coverage_at_5pct_error": 0.225,
1329
- "coverage_at_1pct_error": 0.225,
1330
- "aurc": 0.10718612602417658,
1331
- "error_rate_at_0_9": 0.2,
1332
- "confidence_bias": 0.09523062775922997,
1333
- "top_bins": {
1334
- "0.9": {
1335
- "n": 30,
1336
- "errors": 6,
1337
- "error_rate": 0.2
1338
- },
1339
- "0.95": {
1340
- "n": 26,
1341
- "errors": 3,
1342
- "error_rate": 0.11538461538461539
1343
- },
1344
- "0.99": {
1345
- "n": 15,
1346
- "errors": 1,
1347
- "error_rate": 0.06666666666666667
1348
- }
1349
- },
1350
- "selective": {
1351
- "0.5": {
1352
- "coverage": 0.5,
1353
- "accuracy": 0.9,
1354
- "confidence_cutoff": 0.9774245184450835
1355
- },
1356
- "0.8": {
1357
- "coverage": 0.8,
1358
- "accuracy": 0.8125,
1359
- "confidence_cutoff": 0.8542392185801473
1360
- }
1361
- }
1362
- },
1363
- "sev_swarmtraces_manual_flow": {
1364
- "n": 40,
1365
- "nll": 1.485896746743676,
1366
- "acc": 0.75,
1367
- "ece": 0.21977756775377905,
1368
- "brier": 0.43299049842888415,
1369
- "mean_conf": 0.9214803040956191,
1370
- "confident_error_rate": 0.15,
1371
- "coverage_at_0_9": 0.725,
1372
- "accuracy_at_0_9": 0.7931034482758621,
1373
- "coverage_at_5pct_error": 0.1,
1374
- "coverage_at_1pct_error": 0.1,
1375
- "aurc": 0.14190456225170725,
1376
- "error_rate_at_0_9": 0.20689655172413793,
1377
- "confidence_bias": 0.1714803040956191,
1378
- "top_bins": {
1379
- "0.9": {
1380
- "n": 29,
1381
- "errors": 6,
1382
- "error_rate": 0.20689655172413793
1383
- },
1384
- "0.95": {
1385
- "n": 27,
1386
- "errors": 6,
1387
- "error_rate": 0.2222222222222222
1388
- },
1389
- "0.99": {
1390
- "n": 21,
1391
- "errors": 2,
1392
- "error_rate": 0.09523809523809523
1393
- }
1394
- },
1395
- "selective": {
1396
- "0.5": {
1397
- "coverage": 0.5,
1398
- "accuracy": 0.9,
1399
- "confidence_cutoff": 0.9901066734665328
1400
- },
1401
- "0.8": {
1402
- "coverage": 0.8,
1403
- "accuracy": 0.8125,
1404
- "confidence_cutoff": 0.8291658959545783
1405
- }
1406
- }
1407
- },
1408
- "sev_swarmtraces_manual_prerequisite": {
1409
- "n": 3,
1410
- "nll": 1.0331257533433365e-05,
1411
- "acc": 1.0,
1412
- "ece": 1.0331183196821137e-05,
1413
- "brier": 2.5195789327524225e-10,
1414
- "mean_conf": 0.9999896688168032,
1415
- "confident_error_rate": 0.0,
1416
- "coverage_at_0_9": 1.0,
1417
- "accuracy_at_0_9": 1.0,
1418
- "coverage_at_5pct_error": 1.0,
1419
- "coverage_at_1pct_error": 1.0,
1420
- "aurc": 0.0,
1421
- "error_rate_at_0_9": 0.0,
1422
- "confidence_bias": -1.0331183196821137e-05,
1423
- "top_bins": {
1424
- "0.9": {
1425
- "n": 3,
1426
- "errors": 0,
1427
- "error_rate": 0.0
1428
- },
1429
- "0.95": {
1430
- "n": 3,
1431
- "errors": 0,
1432
- "error_rate": 0.0
1433
- },
1434
- "0.99": {
1435
- "n": 3,
1436
- "errors": 0,
1437
- "error_rate": 0.0
1438
- }
1439
- },
1440
- "selective": {
1441
- "0.5": {
1442
- "coverage": 0.6666666666666666,
1443
- "accuracy": 1.0,
1444
- "confidence_cutoff": 0.9999897312178962
1445
- },
1446
- "0.8": {
1447
- "coverage": 1.0,
1448
- "accuracy": 1.0,
1449
- "confidence_cutoff": 0.9999810314632929
1450
- }
1451
- }
1452
- },
1453
- "tweet_offensive": {
1454
- "n": 8,
1455
- "nll": 0.7172428487540772,
1456
- "acc": 0.75,
1457
- "ece": 0.2179639808956667,
1458
- "brier": 0.44120571426799937,
1459
- "mean_conf": 0.9679639808956667,
1460
- "confident_error_rate": 0.25,
1461
- "coverage_at_0_9": 1.0,
1462
- "accuracy_at_0_9": 0.75,
1463
- "coverage_at_5pct_error": 0.5,
1464
- "coverage_at_1pct_error": 0.5,
1465
- "aurc": 0.09494047619047619,
1466
- "error_rate_at_0_9": 0.25,
1467
- "confidence_bias": 0.2179639808956667,
1468
- "top_bins": {
1469
- "0.9": {
1470
- "n": 8,
1471
- "errors": 2,
1472
- "error_rate": 0.25
1473
- },
1474
- "0.95": {
1475
- "n": 5,
1476
- "errors": 1,
1477
- "error_rate": 0.2
1478
- },
1479
- "0.99": {
1480
- "n": 3,
1481
- "errors": 0,
1482
- "error_rate": 0.0
1483
- }
1484
- },
1485
- "selective": {
1486
- "0.5": {
1487
- "coverage": 0.5,
1488
- "accuracy": 1.0,
1489
- "confidence_cutoff": 0.9897168920827885
1490
- },
1491
- "0.8": {
1492
- "coverage": 0.875,
1493
- "accuracy": 0.8571428571428571,
1494
- "confidence_cutoff": 0.940377631626033
1495
- }
1496
- }
1497
- }
1498
- },
1499
- "variants": {
1500
- "authorized_dns_diagnostic": {
1501
- "n": 4,
1502
- "nll": 0.2732991829356235,
1503
- "acc": 0.75,
1504
- "ece": 0.2014060191596143,
1505
- "brier": 0.1826209474615037,
1506
- "mean_conf": 0.8245034642731555,
1507
- "confident_error_rate": 0.0,
1508
- "coverage_at_0_9": 0.5,
1509
- "accuracy_at_0_9": 1.0,
1510
- "coverage_at_5pct_error": 0.75,
1511
- "coverage_at_1pct_error": 0.75,
1512
- "aurc": 0.0625,
1513
- "error_rate_at_0_9": 0.0,
1514
- "confidence_bias": 0.07450346427315546,
1515
- "top_bins": {
1516
- "0.9": {
1517
- "n": 2,
1518
- "errors": 0,
1519
- "error_rate": 0.0
1520
- },
1521
- "0.95": {
1522
- "n": 2,
1523
- "errors": 0,
1524
- "error_rate": 0.0
1525
- },
1526
- "0.99": {
1527
- "n": 2,
1528
- "errors": 0,
1529
- "error_rate": 0.0
1530
- }
1531
- },
1532
- "selective": {
1533
- "0.5": {
1534
- "coverage": 0.5,
1535
- "accuracy": 1.0,
1536
- "confidence_cutoff": 0.9941628768025105
1537
- },
1538
- "0.8": {
1539
- "coverage": 1.0,
1540
- "accuracy": 0.75,
1541
- "confidence_cutoff": 0.5518189668655396
1542
- }
1543
- }
1544
- },
1545
- "cache_denied_no_reply": {
1546
- "n": 4,
1547
- "nll": 0.6290998127898555,
1548
- "acc": 0.5,
1549
- "ece": 0.19783432109801047,
1550
- "brier": 0.45073491547326267,
1551
- "mean_conf": 0.6842140294824197,
1552
- "confident_error_rate": 0.0,
1553
- "coverage_at_0_9": 0.25,
1554
- "accuracy_at_0_9": 1.0,
1555
- "coverage_at_5pct_error": 0.25,
1556
- "coverage_at_1pct_error": 0.25,
1557
- "aurc": 0.41666666666666663,
1558
- "error_rate_at_0_9": 0.0,
1559
- "confidence_bias": 0.18421402948241972,
1560
- "top_bins": {
1561
- "0.9": {
1562
- "n": 1,
1563
- "errors": 0,
1564
- "error_rate": 0.0
1565
- },
1566
- "0.95": {
1567
- "n": 1,
1568
- "errors": 0,
1569
- "error_rate": 0.0
1570
- },
1571
- "0.99": {
1572
- "n": 0,
1573
- "errors": 0,
1574
- "error_rate": null
1575
- }
1576
- },
1577
- "selective": {
1578
- "0.5": {
1579
- "coverage": 0.5,
1580
- "accuracy": 0.5,
1581
- "confidence_cutoff": 0.5997726159351691
1582
- },
1583
- "0.8": {
1584
- "coverage": 1.0,
1585
- "accuracy": 0.5,
1586
- "confidence_cutoff": 0.5818420334307344
1587
- }
1588
- }
1589
- },
1590
- "clean": {
1591
- "n": 243,
1592
- "nll": 0.6065943824685833,
1593
- "acc": 0.8683127572016461,
1594
- "ece": 0.09195653550207036,
1595
- "brier": 0.23499321102178952,
1596
- "mean_conf": 0.9525644631029061,
1597
- "confident_error_rate": 0.09053497942386832,
1598
- "coverage_at_0_9": 0.8600823045267489,
1599
- "accuracy_at_0_9": 0.8947368421052632,
1600
- "coverage_at_5pct_error": 0.6872427983539094,
1601
- "coverage_at_1pct_error": 0.42386831275720166,
1602
- "aurc": 0.034793809228677,
1603
- "error_rate_at_0_9": 0.10526315789473684,
1604
- "confidence_bias": 0.08425170590125997,
1605
- "top_bins": {
1606
- "0.9": {
1607
- "n": 209,
1608
- "errors": 22,
1609
- "error_rate": 0.10526315789473684
1610
- },
1611
- "0.95": {
1612
- "n": 193,
1613
- "errors": 13,
1614
- "error_rate": 0.06735751295336788
1615
- },
1616
- "0.99": {
1617
- "n": 154,
1618
- "errors": 4,
1619
- "error_rate": 0.025974025974025976
1620
- }
1621
- },
1622
- "selective": {
1623
- "0.5": {
1624
- "coverage": 0.5020576131687243,
1625
- "accuracy": 0.9836065573770492,
1626
- "confidence_cutoff": 0.9985267483650163
1627
- },
1628
- "0.8": {
1629
- "coverage": 0.8024691358024691,
1630
- "accuracy": 0.9230769230769231,
1631
- "confidence_cutoff": 0.946113674977154
1632
- }
1633
- }
1634
- },
1635
- "dns_off_task_reply": {
1636
- "n": 4,
1637
- "nll": 2.500105858228899,
1638
- "acc": 0.25,
1639
- "ece": 0.7354215562262277,
1640
- "brier": 1.1856881009881057,
1641
- "mean_conf": 0.7180991063824902,
1642
- "confident_error_rate": 0.25,
1643
- "coverage_at_0_9": 0.25,
1644
- "accuracy_at_0_9": 0.0,
1645
- "coverage_at_5pct_error": 0.0,
1646
- "coverage_at_1pct_error": 0.0,
1647
- "aurc": 0.9375,
1648
- "error_rate_at_0_9": 1.0,
1649
- "confidence_bias": 0.4680991063824902,
1650
- "top_bins": {
1651
- "0.9": {
1652
- "n": 1,
1653
- "errors": 1,
1654
- "error_rate": 1.0
1655
- },
1656
- "0.95": {
1657
- "n": 1,
1658
- "errors": 1,
1659
- "error_rate": 1.0
1660
- },
1661
- "0.99": {
1662
- "n": 1,
1663
- "errors": 1,
1664
- "error_rate": 1.0
1665
- }
1666
- },
1667
- "selective": {
1668
- "0.5": {
1669
- "coverage": 0.5,
1670
- "accuracy": 0.0,
1671
- "confidence_cutoff": 0.7394568794916744
1672
- },
1673
- "0.8": {
1674
- "coverage": 1.0,
1675
- "accuracy": 0.25,
1676
- "confidence_cutoff": 0.465355100312525
1677
- }
1678
- }
1679
- },
1680
- "dns_unhelpful_reply": {
1681
- "n": 4,
1682
- "nll": 0.6130712887407269,
1683
- "acc": 0.75,
1684
- "ece": 0.2694959031277442,
1685
- "brier": 0.34519095756377294,
1686
- "mean_conf": 0.7624970557751488,
1687
- "confident_error_rate": 0.0,
1688
- "coverage_at_0_9": 0.5,
1689
- "accuracy_at_0_9": 1.0,
1690
- "coverage_at_5pct_error": 0.75,
1691
- "coverage_at_1pct_error": 0.75,
1692
- "aurc": 0.0625,
1693
- "error_rate_at_0_9": 0.0,
1694
- "confidence_bias": 0.012497055775148813,
1695
- "top_bins": {
1696
- "0.9": {
1697
- "n": 2,
1698
- "errors": 0,
1699
- "error_rate": 0.0
1700
- },
1701
- "0.95": {
1702
- "n": 1,
1703
- "errors": 0,
1704
- "error_rate": 0.0
1705
- },
1706
- "0.99": {
1707
- "n": 0,
1708
- "errors": 0,
1709
- "error_rate": null
1710
- }
1711
- },
1712
- "selective": {
1713
- "0.5": {
1714
- "coverage": 0.5,
1715
- "accuracy": 1.0,
1716
- "confidence_cutoff": 0.9232891824527384
1717
- },
1718
- "0.8": {
1719
- "coverage": 1.0,
1720
- "accuracy": 0.75,
1721
- "confidence_cutoff": 0.563985917805786
1722
- }
1723
- }
1724
- },
1725
- "dns_without_reply": {
1726
- "n": 4,
1727
- "nll": 0.6878185272523656,
1728
- "acc": 0.5,
1729
- "ece": 0.20221082782282854,
1730
- "brier": 0.5043315891024525,
1731
- "mean_conf": 0.7000800429377261,
1732
- "confident_error_rate": 0.0,
1733
- "coverage_at_0_9": 0.25,
1734
- "accuracy_at_0_9": 1.0,
1735
- "coverage_at_5pct_error": 0.25,
1736
- "coverage_at_1pct_error": 0.25,
1737
- "aurc": 0.41666666666666663,
1738
- "error_rate_at_0_9": 0.0,
1739
- "confidence_bias": 0.20008004293772608,
1740
- "top_bins": {
1741
- "0.9": {
1742
- "n": 1,
1743
- "errors": 0,
1744
- "error_rate": 0.0
1745
- },
1746
- "0.95": {
1747
- "n": 1,
1748
- "errors": 0,
1749
- "error_rate": 0.0
1750
- },
1751
- "0.99": {
1752
- "n": 1,
1753
- "errors": 0,
1754
- "error_rate": 0.0
1755
- }
1756
- },
1757
- "selective": {
1758
- "0.5": {
1759
- "coverage": 0.5,
1760
- "accuracy": 0.5,
1761
- "confidence_cutoff": 0.6943924514728984
1762
- },
1763
- "0.8": {
1764
- "coverage": 1.0,
1765
- "accuracy": 0.5,
1766
- "confidence_cutoff": 0.5490827128017757
1767
- }
1768
- }
1769
- },
1770
- "none_absent": {
1771
- "n": 3,
1772
- "nll": 1.3777335821359413,
1773
- "acc": 0.6666666666666666,
1774
- "ece": 0.24885046106665176,
1775
- "brier": 0.5885968900899531,
1776
- "mean_conf": 0.9151737396988762,
1777
- "confident_error_rate": 0.0,
1778
- "coverage_at_0_9": 0.3333333333333333,
1779
- "accuracy_at_0_9": 1.0,
1780
- "coverage_at_5pct_error": 0.6666666666666666,
1781
- "coverage_at_1pct_error": 0.6666666666666666,
1782
- "aurc": 0.1111111111111111,
1783
- "error_rate_at_0_9": 0.0,
1784
- "confidence_bias": 0.24850707303220954,
1785
- "top_bins": {
1786
- "0.9": {
1787
- "n": 1,
1788
- "errors": 0,
1789
- "error_rate": 0.0
1790
- },
1791
- "0.95": {
1792
- "n": 1,
1793
- "errors": 0,
1794
- "error_rate": 0.0
1795
- },
1796
- "0.99": {
1797
- "n": 1,
1798
- "errors": 0,
1799
- "error_rate": 0.0
1800
- }
1801
- },
1802
- "selective": {
1803
- "0.5": {
1804
- "coverage": 0.6666666666666666,
1805
- "accuracy": 1.0,
1806
- "confidence_cutoff": 0.8739663040318306
1807
- },
1808
- "0.8": {
1809
- "coverage": 1.0,
1810
- "accuracy": 0.6666666666666666,
1811
- "confidence_cutoff": 0.8720699971164613
1812
- }
1813
- }
1814
- },
1815
- "none_present": {
1816
- "n": 3,
1817
- "nll": 3.3971590445247757,
1818
- "acc": 0.0,
1819
- "ece": 0.806803931885096,
1820
- "brier": 1.589050038207996,
1821
- "mean_conf": 0.8068039318850961,
1822
- "confident_error_rate": 0.3333333333333333,
1823
- "coverage_at_0_9": 0.3333333333333333,
1824
- "accuracy_at_0_9": 0.0,
1825
- "coverage_at_5pct_error": 0.0,
1826
- "coverage_at_1pct_error": 0.0,
1827
- "aurc": 1.0,
1828
- "error_rate_at_0_9": 1.0,
1829
- "confidence_bias": 0.8068039318850961,
1830
- "top_bins": {
1831
- "0.9": {
1832
- "n": 1,
1833
- "errors": 1,
1834
- "error_rate": 1.0
1835
- },
1836
- "0.95": {
1837
- "n": 1,
1838
- "errors": 1,
1839
- "error_rate": 1.0
1840
- },
1841
- "0.99": {
1842
- "n": 0,
1843
- "errors": 0,
1844
- "error_rate": null
1845
- }
1846
- },
1847
- "selective": {
1848
- "0.5": {
1849
- "coverage": 0.6666666666666666,
1850
- "accuracy": 0.0,
1851
- "confidence_cutoff": 0.8592945698759779
1852
- },
1853
- "0.8": {
1854
- "coverage": 1.0,
1855
- "accuracy": 0.0,
1856
- "confidence_cutoff": 0.5949657569957464
1857
- }
1858
- }
1859
- },
1860
- "operator_terminated": {
1861
- "n": 4,
1862
- "nll": 1.0359158594833722,
1863
- "acc": 0.75,
1864
- "ece": 0.2925896655492872,
1865
- "brier": 0.4933779585120464,
1866
- "mean_conf": 0.9106235817198012,
1867
- "confident_error_rate": 0.25,
1868
- "coverage_at_0_9": 0.75,
1869
- "accuracy_at_0_9": 0.6666666666666666,
1870
- "coverage_at_5pct_error": 0.5,
1871
- "coverage_at_1pct_error": 0.5,
1872
- "aurc": 0.14583333333333331,
1873
- "error_rate_at_0_9": 0.3333333333333333,
1874
- "confidence_bias": 0.16062358171980118,
1875
- "top_bins": {
1876
- "0.9": {
1877
- "n": 3,
1878
- "errors": 1,
1879
- "error_rate": 0.3333333333333333
1880
- },
1881
- "0.95": {
1882
- "n": 2,
1883
- "errors": 0,
1884
- "error_rate": 0.0
1885
- },
1886
- "0.99": {
1887
- "n": 1,
1888
- "errors": 0,
1889
- "error_rate": 0.0
1890
- }
1891
- },
1892
- "selective": {
1893
- "0.5": {
1894
- "coverage": 0.5,
1895
- "accuracy": 1.0,
1896
- "confidence_cutoff": 0.9713811910576418
1897
- },
1898
- "0.8": {
1899
- "coverage": 1.0,
1900
- "accuracy": 0.75,
1901
- "confidence_cutoff": 0.736067832341028
1902
- }
1903
- }
1904
- },
1905
- "ordinary_resolution": {
1906
- "n": 4,
1907
- "nll": 0.04811475260764305,
1908
- "acc": 1.0,
1909
- "ece": 0.04485818654799875,
1910
- "brier": 0.011362447845748204,
1911
- "mean_conf": 0.9551418134520012,
1912
- "confident_error_rate": 0.0,
1913
- "coverage_at_0_9": 0.75,
1914
- "accuracy_at_0_9": 1.0,
1915
- "coverage_at_5pct_error": 1.0,
1916
- "coverage_at_1pct_error": 1.0,
1917
- "aurc": 0.0,
1918
- "error_rate_at_0_9": 0.0,
1919
- "confidence_bias": -0.04485818654799878,
1920
- "top_bins": {
1921
- "0.9": {
1922
- "n": 3,
1923
- "errors": 0,
1924
- "error_rate": 0.0
1925
- },
1926
- "0.95": {
1927
- "n": 3,
1928
- "errors": 0,
1929
- "error_rate": 0.0
1930
- },
1931
- "0.99": {
1932
- "n": 2,
1933
- "errors": 0,
1934
- "error_rate": 0.0
1935
- }
1936
- },
1937
- "selective": {
1938
- "0.5": {
1939
- "coverage": 0.5,
1940
- "accuracy": 1.0,
1941
- "confidence_cutoff": 0.9973615074301722
1942
- },
1943
- "0.8": {
1944
- "coverage": 1.0,
1945
- "accuracy": 1.0,
1946
- "confidence_cutoff": 0.848997818390719
1947
- }
1948
- }
1949
- },
1950
- "permuted": {
1951
- "n": 3,
1952
- "nll": 1.5849320188552285,
1953
- "acc": 0.3333333333333333,
1954
- "ece": 0.5115039138500523,
1955
- "brier": 0.948500922906351,
1956
- "mean_conf": 0.834218301404813,
1957
- "confident_error_rate": 0.0,
1958
- "coverage_at_0_9": 0.3333333333333333,
1959
- "accuracy_at_0_9": 1.0,
1960
- "coverage_at_5pct_error": 0.3333333333333333,
1961
- "coverage_at_1pct_error": 0.3333333333333333,
1962
- "aurc": 0.38888888888888884,
1963
- "error_rate_at_0_9": 0.0,
1964
- "confidence_bias": 0.5008849680714798,
1965
- "top_bins": {
1966
- "0.9": {
1967
- "n": 1,
1968
- "errors": 0,
1969
- "error_rate": 0.0
1970
- },
1971
- "0.95": {
1972
- "n": 1,
1973
- "errors": 0,
1974
- "error_rate": 0.0
1975
- },
1976
- "0.99": {
1977
- "n": 0,
1978
- "errors": 0,
1979
- "error_rate": null
1980
- }
1981
- },
1982
- "selective": {
1983
- "0.5": {
1984
- "coverage": 0.6666666666666666,
1985
- "accuracy": 0.5,
1986
- "confidence_cutoff": 0.8701283299219073
1987
- },
1988
- "0.8": {
1989
- "coverage": 1.0,
1990
- "accuracy": 0.3333333333333333,
1991
- "confidence_cutoff": 0.6484549929603911
1992
- }
1993
- }
1994
- },
1995
- "review_acknowledged": {
1996
- "n": 4,
1997
- "nll": 0.9430886337427746,
1998
- "acc": 0.75,
1999
- "ece": 0.26428833978133903,
2000
- "brier": 0.46621510579801406,
2001
- "mean_conf": 0.9076209540563065,
2002
- "confident_error_rate": 0.25,
2003
- "coverage_at_0_9": 0.75,
2004
- "accuracy_at_0_9": 0.6666666666666666,
2005
- "coverage_at_5pct_error": 0.5,
2006
- "coverage_at_1pct_error": 0.5,
2007
- "aurc": 0.14583333333333331,
2008
- "error_rate_at_0_9": 0.3333333333333333,
2009
- "confidence_bias": 0.1576209540563065,
2010
- "top_bins": {
2011
- "0.9": {
2012
- "n": 3,
2013
- "errors": 1,
2014
- "error_rate": 0.3333333333333333
2015
- },
2016
- "0.95": {
2017
- "n": 1,
2018
- "errors": 0,
2019
- "error_rate": 0.0
2020
- },
2021
- "0.99": {
2022
- "n": 1,
2023
- "errors": 0,
2024
- "error_rate": 0.0
2025
- }
2026
- },
2027
- "selective": {
2028
- "0.5": {
2029
- "coverage": 0.5,
2030
- "accuracy": 1.0,
2031
- "confidence_cutoff": 0.9371813046912169
2032
- },
2033
- "0.8": {
2034
- "coverage": 1.0,
2035
- "accuracy": 0.75,
2036
- "confidence_cutoff": 0.786665228549935
2037
- }
2038
- }
2039
- }
2040
- },
2041
- "heldout_tasks": {
2042
- "composition_held_and_or": {
2043
- "n": 32,
2044
- "nll": 0.24171629359001318,
2045
- "acc": 0.90625,
2046
- "ece": 0.08267706695037225,
2047
- "brier": 0.1580607438572586,
2048
- "mean_conf": 0.9889270669503722,
2049
- "confident_error_rate": 0.09375,
2050
- "coverage_at_0_9": 1.0,
2051
- "accuracy_at_0_9": 0.90625,
2052
- "coverage_at_5pct_error": 0.9375,
2053
- "coverage_at_1pct_error": 0.90625,
2054
- "aurc": 0.0069010416666666664,
2055
- "error_rate_at_0_9": 0.09375,
2056
- "confidence_bias": 0.08267706695037225,
2057
- "top_bins": {
2058
- "0.9": {
2059
- "n": 32,
2060
- "errors": 3,
2061
- "error_rate": 0.09375
2062
- },
2063
- "0.95": {
2064
- "n": 30,
2065
- "errors": 1,
2066
- "error_rate": 0.03333333333333333
2067
- },
2068
- "0.99": {
2069
- "n": 26,
2070
- "errors": 0,
2071
- "error_rate": 0.0
2072
- }
2073
- },
2074
- "selective": {
2075
- "0.5": {
2076
- "coverage": 0.5,
2077
- "accuracy": 1.0,
2078
- "confidence_cutoff": 0.9999364856833449
2079
- },
2080
- "0.8": {
2081
- "coverage": 0.8125,
2082
- "accuracy": 1.0,
2083
- "confidence_cutoff": 0.9996934954887139
2084
- }
2085
- }
2086
- },
2087
- "contrastive_authorization": {
2088
- "n": 16,
2089
- "nll": 1.502653141036028e-05,
2090
- "acc": 1.0,
2091
- "ece": 1.50264042420023e-05,
2092
- "brier": 5.139828295030656e-10,
2093
- "mean_conf": 0.999984973595758,
2094
- "confident_error_rate": 0.0,
2095
- "coverage_at_0_9": 1.0,
2096
- "accuracy_at_0_9": 1.0,
2097
- "coverage_at_5pct_error": 1.0,
2098
- "coverage_at_1pct_error": 1.0,
2099
- "aurc": 0.0,
2100
- "error_rate_at_0_9": 0.0,
2101
- "confidence_bias": -1.50264042420023e-05,
2102
- "top_bins": {
2103
- "0.9": {
2104
- "n": 16,
2105
- "errors": 0,
2106
- "error_rate": 0.0
2107
- },
2108
- "0.95": {
2109
- "n": 16,
2110
- "errors": 0,
2111
- "error_rate": 0.0
2112
- },
2113
- "0.99": {
2114
- "n": 16,
2115
- "errors": 0,
2116
- "error_rate": 0.0
2117
- }
2118
- },
2119
- "selective": {
2120
- "0.5": {
2121
- "coverage": 0.5,
2122
- "accuracy": 1.0,
2123
- "confidence_cutoff": 0.9999862890472545
2124
- },
2125
- "0.8": {
2126
- "coverage": 0.8125,
2127
- "accuracy": 1.0,
2128
- "confidence_cutoff": 0.9999813198967473
2129
- }
2130
- }
2131
- },
2132
- "emotion": {
2133
- "n": 8,
2134
- "nll": 1.4194273414243654,
2135
- "acc": 0.625,
2136
- "ece": 0.3420785503418361,
2137
- "brier": 0.6845960520300963,
2138
- "mean_conf": 0.967078550341836,
2139
- "confident_error_rate": 0.25,
2140
- "coverage_at_0_9": 0.875,
2141
- "accuracy_at_0_9": 0.7142857142857143,
2142
- "coverage_at_5pct_error": 0.5,
2143
- "coverage_at_1pct_error": 0.5,
2144
- "aurc": 0.12842261904761904,
2145
- "error_rate_at_0_9": 0.2857142857142857,
2146
- "confidence_bias": 0.342078550341836,
2147
- "top_bins": {
2148
- "0.9": {
2149
- "n": 7,
2150
- "errors": 2,
2151
- "error_rate": 0.2857142857142857
2152
- },
2153
- "0.95": {
2154
- "n": 5,
2155
- "errors": 1,
2156
- "error_rate": 0.2
2157
- },
2158
- "0.99": {
2159
- "n": 3,
2160
- "errors": 0,
2161
- "error_rate": 0.0
2162
- }
2163
- },
2164
- "selective": {
2165
- "0.5": {
2166
- "coverage": 0.5,
2167
- "accuracy": 1.0,
2168
- "confidence_cutoff": 0.9869563807629563
2169
- },
2170
- "0.8": {
2171
- "coverage": 0.875,
2172
- "accuracy": 0.7142857142857143,
2173
- "confidence_cutoff": 0.936063568486461
2174
- }
2175
- }
2176
- },
2177
- "mmlu": {
2178
- "n": 8,
2179
- "nll": 1.3129749769552814,
2180
- "acc": 0.625,
2181
- "ece": 0.15765322690405312,
2182
- "brier": 0.6484351711065982,
2183
- "mean_conf": 0.7567759888850003,
2184
- "confident_error_rate": 0.125,
2185
- "coverage_at_0_9": 0.5,
2186
- "accuracy_at_0_9": 0.75,
2187
- "coverage_at_5pct_error": 0.375,
2188
- "coverage_at_1pct_error": 0.375,
2189
- "aurc": 0.22336309523809525,
2190
- "error_rate_at_0_9": 0.25,
2191
- "confidence_bias": 0.1317759888850003,
2192
- "top_bins": {
2193
- "0.9": {
2194
- "n": 4,
2195
- "errors": 1,
2196
- "error_rate": 0.25
2197
- },
2198
- "0.95": {
2199
- "n": 3,
2200
- "errors": 0,
2201
- "error_rate": 0.0
2202
- },
2203
- "0.99": {
2204
- "n": 1,
2205
- "errors": 0,
2206
- "error_rate": 0.0
2207
- }
2208
- },
2209
- "selective": {
2210
- "0.5": {
2211
- "coverage": 0.5,
2212
- "accuracy": 0.75,
2213
- "confidence_cutoff": 0.9327223999584052
2214
- },
2215
- "0.8": {
2216
- "coverage": 0.875,
2217
- "accuracy": 0.5714285714285714,
2218
- "confidence_cutoff": 0.485476400811401
2219
- }
2220
- }
2221
- },
2222
- "otrf_sysmon_file": {
2223
- "n": 16,
2224
- "nll": 0.008605175298168915,
2225
- "acc": 1.0,
2226
- "ece": 0.008437849080092708,
2227
- "brier": 0.0005211762382657875,
2228
- "mean_conf": 0.9915621509199073,
2229
- "confident_error_rate": 0.0,
2230
- "coverage_at_0_9": 1.0,
2231
- "accuracy_at_0_9": 1.0,
2232
- "coverage_at_5pct_error": 1.0,
2233
- "coverage_at_1pct_error": 1.0,
2234
- "aurc": 0.0,
2235
- "error_rate_at_0_9": 0.0,
2236
- "confidence_bias": -0.008437849080092708,
2237
- "top_bins": {
2238
- "0.9": {
2239
- "n": 16,
2240
- "errors": 0,
2241
- "error_rate": 0.0
2242
- },
2243
- "0.95": {
2244
- "n": 15,
2245
- "errors": 0,
2246
- "error_rate": 0.0
2247
- },
2248
- "0.99": {
2249
- "n": 13,
2250
- "errors": 0,
2251
- "error_rate": 0.0
2252
- }
2253
- },
2254
- "selective": {
2255
- "0.5": {
2256
- "coverage": 0.5,
2257
- "accuracy": 1.0,
2258
- "confidence_cutoff": 0.9989837483137961
2259
- },
2260
- "0.8": {
2261
- "coverage": 0.8125,
2262
- "accuracy": 1.0,
2263
- "confidence_cutoff": 0.9925042879134804
2264
- }
2265
- }
2266
- },
2267
- "otrf_sysmon_join": {
2268
- "n": 16,
2269
- "nll": 0.12967885980971613,
2270
- "acc": 0.9375,
2271
- "ece": 0.0857233928404766,
2272
- "brier": 0.08146985002971384,
2273
- "mean_conf": 0.946506108546294,
2274
- "confident_error_rate": 0.0,
2275
- "coverage_at_0_9": 0.6875,
2276
- "accuracy_at_0_9": 1.0,
2277
- "coverage_at_5pct_error": 0.9375,
2278
- "coverage_at_1pct_error": 0.9375,
2279
- "aurc": 0.00390625,
2280
- "error_rate_at_0_9": 0.0,
2281
- "confidence_bias": 0.009006108546294023,
2282
- "top_bins": {
2283
- "0.9": {
2284
- "n": 11,
2285
- "errors": 0,
2286
- "error_rate": 0.0
2287
- },
2288
- "0.95": {
2289
- "n": 11,
2290
- "errors": 0,
2291
- "error_rate": 0.0
2292
- },
2293
- "0.99": {
2294
- "n": 9,
2295
- "errors": 0,
2296
- "error_rate": 0.0
2297
- }
2298
- },
2299
- "selective": {
2300
- "0.5": {
2301
- "coverage": 0.5,
2302
- "accuracy": 1.0,
2303
- "confidence_cutoff": 0.9970209747654696
2304
- },
2305
- "0.8": {
2306
- "coverage": 0.8125,
2307
- "accuracy": 1.0,
2308
- "confidence_cutoff": 0.8702734279589676
2309
- }
2310
- }
2311
- },
2312
- "otrf_sysmon_process": {
2313
- "n": 24,
2314
- "nll": 0.16156109121321785,
2315
- "acc": 0.9583333333333334,
2316
- "ece": 0.05273169620250518,
2317
- "brier": 0.0899182770612567,
2318
- "mean_conf": 0.9768730534919899,
2319
- "confident_error_rate": 0.041666666666666664,
2320
- "coverage_at_0_9": 0.9583333333333334,
2321
- "accuracy_at_0_9": 0.9565217391304348,
2322
- "coverage_at_5pct_error": 1.0,
2323
- "coverage_at_1pct_error": 0.9166666666666666,
2324
- "aurc": 0.0035477053140096615,
2325
- "error_rate_at_0_9": 0.043478260869565216,
2326
- "confidence_bias": 0.01853972015865657,
2327
- "top_bins": {
2328
- "0.9": {
2329
- "n": 23,
2330
- "errors": 1,
2331
- "error_rate": 0.043478260869565216
2332
- },
2333
- "0.95": {
2334
- "n": 22,
2335
- "errors": 0,
2336
- "error_rate": 0.0
2337
- },
2338
- "0.99": {
2339
- "n": 19,
2340
- "errors": 0,
2341
- "error_rate": 0.0
2342
- }
2343
- },
2344
- "selective": {
2345
- "0.5": {
2346
- "coverage": 0.5,
2347
- "accuracy": 1.0,
2348
- "confidence_cutoff": 0.9994396085787187
2349
- },
2350
- "0.8": {
2351
- "coverage": 0.8333333333333334,
2352
- "accuracy": 1.0,
2353
- "confidence_cutoff": 0.9832554759400715
2354
- }
2355
- }
2356
- },
2357
- "otrf_sysmon_registry": {
2358
- "n": 8,
2359
- "nll": 0.010723606364118954,
2360
- "acc": 1.0,
2361
- "ece": 0.010542385937563847,
2362
- "brier": 0.0006959554565011166,
2363
- "mean_conf": 0.9894576140624362,
2364
- "confident_error_rate": 0.0,
2365
- "coverage_at_0_9": 1.0,
2366
- "accuracy_at_0_9": 1.0,
2367
- "coverage_at_5pct_error": 1.0,
2368
- "coverage_at_1pct_error": 1.0,
2369
- "aurc": 0.0,
2370
- "error_rate_at_0_9": 0.0,
2371
- "confidence_bias": -0.010542385937563847,
2372
- "top_bins": {
2373
- "0.9": {
2374
- "n": 8,
2375
- "errors": 0,
2376
- "error_rate": 0.0
2377
- },
2378
- "0.95": {
2379
- "n": 8,
2380
- "errors": 0,
2381
- "error_rate": 0.0
2382
- },
2383
- "0.99": {
2384
- "n": 5,
2385
- "errors": 0,
2386
- "error_rate": 0.0
2387
- }
2388
- },
2389
- "selective": {
2390
- "0.5": {
2391
- "coverage": 0.5,
2392
- "accuracy": 1.0,
2393
- "confidence_cutoff": 0.9972651302471073
2394
- },
2395
- "0.8": {
2396
- "coverage": 0.875,
2397
- "accuracy": 1.0,
2398
- "confidence_cutoff": 0.9840692701748682
2399
- }
2400
- }
2401
- },
2402
- "paws": {
2403
- "n": 8,
2404
- "nll": 1.0368545966100273,
2405
- "acc": 0.75,
2406
- "ece": 0.22498828525853157,
2407
- "brier": 0.34163600039007924,
2408
- "mean_conf": 0.9122532529823204,
2409
- "confident_error_rate": 0.125,
2410
- "coverage_at_0_9": 0.75,
2411
- "accuracy_at_0_9": 0.8333333333333334,
2412
- "coverage_at_5pct_error": 0.5,
2413
- "coverage_at_1pct_error": 0.5,
2414
- "aurc": 0.09494047619047619,
2415
- "error_rate_at_0_9": 0.16666666666666666,
2416
- "confidence_bias": 0.1622532529823204,
2417
- "top_bins": {
2418
- "0.9": {
2419
- "n": 6,
2420
- "errors": 1,
2421
- "error_rate": 0.16666666666666666
2422
- },
2423
- "0.95": {
2424
- "n": 6,
2425
- "errors": 1,
2426
- "error_rate": 0.16666666666666666
2427
- },
2428
- "0.99": {
2429
- "n": 6,
2430
- "errors": 1,
2431
- "error_rate": 0.16666666666666666
2432
- }
2433
- },
2434
- "selective": {
2435
- "0.5": {
2436
- "coverage": 0.5,
2437
- "accuracy": 1.0,
2438
- "confidence_cutoff": 0.9992691391151949
2439
- },
2440
- "0.8": {
2441
- "coverage": 0.875,
2442
- "accuracy": 0.8571428571428571,
2443
- "confidence_cutoff": 0.749059870895155
2444
- }
2445
- }
2446
- },
2447
- "qnli": {
2448
- "n": 8,
2449
- "nll": 0.005181466224742151,
2450
- "acc": 1.0,
2451
- "ece": 0.005129334884136472,
2452
- "brier": 0.00020524982613720548,
2453
- "mean_conf": 0.9948706651158635,
2454
- "confident_error_rate": 0.0,
2455
- "coverage_at_0_9": 1.0,
2456
- "accuracy_at_0_9": 1.0,
2457
- "coverage_at_5pct_error": 1.0,
2458
- "coverage_at_1pct_error": 1.0,
2459
- "aurc": 0.0,
2460
- "error_rate_at_0_9": 0.0,
2461
- "confidence_bias": -0.005129334884136472,
2462
- "top_bins": {
2463
- "0.9": {
2464
- "n": 8,
2465
- "errors": 0,
2466
- "error_rate": 0.0
2467
- },
2468
- "0.95": {
2469
- "n": 8,
2470
- "errors": 0,
2471
- "error_rate": 0.0
2472
- },
2473
- "0.99": {
2474
- "n": 6,
2475
- "errors": 0,
2476
- "error_rate": 0.0
2477
- }
2478
- },
2479
- "selective": {
2480
- "0.5": {
2481
- "coverage": 0.5,
2482
- "accuracy": 1.0,
2483
- "confidence_cutoff": 0.999952371295075
2484
- },
2485
- "0.8": {
2486
- "coverage": 0.875,
2487
- "accuracy": 1.0,
2488
- "confidence_cutoff": 0.98872862957213
2489
- }
2490
- }
2491
- },
2492
- "sciq": {
2493
- "n": 8,
2494
- "nll": 0.00023928943016782004,
2495
- "acc": 1.0,
2496
- "ece": 0.0002391650142747892,
2497
- "brier": 4.955209174865095e-07,
2498
- "mean_conf": 0.9997608349857252,
2499
- "confident_error_rate": 0.0,
2500
- "coverage_at_0_9": 1.0,
2501
- "accuracy_at_0_9": 1.0,
2502
- "coverage_at_5pct_error": 1.0,
2503
- "coverage_at_1pct_error": 1.0,
2504
- "aurc": 0.0,
2505
- "error_rate_at_0_9": 0.0,
2506
- "confidence_bias": -0.0002391650142747892,
2507
- "top_bins": {
2508
- "0.9": {
2509
- "n": 8,
2510
- "errors": 0,
2511
- "error_rate": 0.0
2512
- },
2513
- "0.95": {
2514
- "n": 8,
2515
- "errors": 0,
2516
- "error_rate": 0.0
2517
- },
2518
- "0.99": {
2519
- "n": 8,
2520
- "errors": 0,
2521
- "error_rate": 0.0
2522
- }
2523
- },
2524
- "selective": {
2525
- "0.5": {
2526
- "coverage": 0.5,
2527
- "accuracy": 1.0,
2528
- "confidence_cutoff": 0.9999911919978735
2529
- },
2530
- "0.8": {
2531
- "coverage": 0.875,
2532
- "accuracy": 1.0,
2533
- "confidence_cutoff": 0.9994259298933852
2534
- }
2535
- }
2536
- },
2537
- "sev_swarmtraces_manual_credential": {
2538
- "n": 40,
2539
- "nll": 0.9530052125004369,
2540
- "acc": 0.825,
2541
- "ece": 0.17316335764162732,
2542
- "brier": 0.3580423587545788,
2543
- "mean_conf": 0.9202306277592299,
2544
- "confident_error_rate": 0.15,
2545
- "coverage_at_0_9": 0.75,
2546
- "accuracy_at_0_9": 0.8,
2547
- "coverage_at_5pct_error": 0.225,
2548
- "coverage_at_1pct_error": 0.225,
2549
- "aurc": 0.10718612602417658,
2550
- "error_rate_at_0_9": 0.2,
2551
- "confidence_bias": 0.09523062775922997,
2552
- "top_bins": {
2553
- "0.9": {
2554
- "n": 30,
2555
- "errors": 6,
2556
- "error_rate": 0.2
2557
- },
2558
- "0.95": {
2559
- "n": 26,
2560
- "errors": 3,
2561
- "error_rate": 0.11538461538461539
2562
- },
2563
- "0.99": {
2564
- "n": 15,
2565
- "errors": 1,
2566
- "error_rate": 0.06666666666666667
2567
- }
2568
- },
2569
- "selective": {
2570
- "0.5": {
2571
- "coverage": 0.5,
2572
- "accuracy": 0.9,
2573
- "confidence_cutoff": 0.9774245184450835
2574
- },
2575
- "0.8": {
2576
- "coverage": 0.8,
2577
- "accuracy": 0.8125,
2578
- "confidence_cutoff": 0.8542392185801473
2579
- }
2580
- }
2581
- },
2582
- "sev_swarmtraces_manual_flow": {
2583
- "n": 40,
2584
- "nll": 1.485896746743676,
2585
- "acc": 0.75,
2586
- "ece": 0.21977756775377905,
2587
- "brier": 0.43299049842888415,
2588
- "mean_conf": 0.9214803040956191,
2589
- "confident_error_rate": 0.15,
2590
- "coverage_at_0_9": 0.725,
2591
- "accuracy_at_0_9": 0.7931034482758621,
2592
- "coverage_at_5pct_error": 0.1,
2593
- "coverage_at_1pct_error": 0.1,
2594
- "aurc": 0.14190456225170725,
2595
- "error_rate_at_0_9": 0.20689655172413793,
2596
- "confidence_bias": 0.1714803040956191,
2597
- "top_bins": {
2598
- "0.9": {
2599
- "n": 29,
2600
- "errors": 6,
2601
- "error_rate": 0.20689655172413793
2602
- },
2603
- "0.95": {
2604
- "n": 27,
2605
- "errors": 6,
2606
- "error_rate": 0.2222222222222222
2607
- },
2608
- "0.99": {
2609
- "n": 21,
2610
- "errors": 2,
2611
- "error_rate": 0.09523809523809523
2612
- }
2613
- },
2614
- "selective": {
2615
- "0.5": {
2616
- "coverage": 0.5,
2617
- "accuracy": 0.9,
2618
- "confidence_cutoff": 0.9901066734665328
2619
- },
2620
- "0.8": {
2621
- "coverage": 0.8,
2622
- "accuracy": 0.8125,
2623
- "confidence_cutoff": 0.8291658959545783
2624
- }
2625
- }
2626
- },
2627
- "sev_swarmtraces_manual_prerequisite": {
2628
- "n": 3,
2629
- "nll": 1.0331257533433365e-05,
2630
- "acc": 1.0,
2631
- "ece": 1.0331183196821137e-05,
2632
- "brier": 2.5195789327524225e-10,
2633
- "mean_conf": 0.9999896688168032,
2634
- "confident_error_rate": 0.0,
2635
- "coverage_at_0_9": 1.0,
2636
- "accuracy_at_0_9": 1.0,
2637
- "coverage_at_5pct_error": 1.0,
2638
- "coverage_at_1pct_error": 1.0,
2639
- "aurc": 0.0,
2640
- "error_rate_at_0_9": 0.0,
2641
- "confidence_bias": -1.0331183196821137e-05,
2642
- "top_bins": {
2643
- "0.9": {
2644
- "n": 3,
2645
- "errors": 0,
2646
- "error_rate": 0.0
2647
- },
2648
- "0.95": {
2649
- "n": 3,
2650
- "errors": 0,
2651
- "error_rate": 0.0
2652
- },
2653
- "0.99": {
2654
- "n": 3,
2655
- "errors": 0,
2656
- "error_rate": 0.0
2657
- }
2658
- },
2659
- "selective": {
2660
- "0.5": {
2661
- "coverage": 0.6666666666666666,
2662
- "accuracy": 1.0,
2663
- "confidence_cutoff": 0.9999897312178962
2664
- },
2665
- "0.8": {
2666
- "coverage": 1.0,
2667
- "accuracy": 1.0,
2668
- "confidence_cutoff": 0.9999810314632929
2669
- }
2670
- }
2671
- },
2672
- "tweet_offensive": {
2673
- "n": 8,
2674
- "nll": 0.7172428487540772,
2675
- "acc": 0.75,
2676
- "ece": 0.2179639808956667,
2677
- "brier": 0.44120571426799937,
2678
- "mean_conf": 0.9679639808956667,
2679
- "confident_error_rate": 0.25,
2680
- "coverage_at_0_9": 1.0,
2681
- "accuracy_at_0_9": 0.75,
2682
- "coverage_at_5pct_error": 0.5,
2683
- "coverage_at_1pct_error": 0.5,
2684
- "aurc": 0.09494047619047619,
2685
- "error_rate_at_0_9": 0.25,
2686
- "confidence_bias": 0.2179639808956667,
2687
- "top_bins": {
2688
- "0.9": {
2689
- "n": 8,
2690
- "errors": 2,
2691
- "error_rate": 0.25
2692
- },
2693
- "0.95": {
2694
- "n": 5,
2695
- "errors": 1,
2696
- "error_rate": 0.2
2697
- },
2698
- "0.99": {
2699
- "n": 3,
2700
- "errors": 0,
2701
- "error_rate": 0.0
2702
- }
2703
- },
2704
- "selective": {
2705
- "0.5": {
2706
- "coverage": 0.5,
2707
- "accuracy": 1.0,
2708
- "confidence_cutoff": 0.9897168920827885
2709
- },
2710
- "0.8": {
2711
- "coverage": 0.875,
2712
- "accuracy": 0.8571428571428571,
2713
- "confidence_cutoff": 0.940377631626033
2714
- }
2715
- }
2716
- }
2717
- },
2718
- "permutation": {
2719
- "n": 3,
2720
- "mean_max_delta": 0.04297798881434687,
2721
- "flip_rate": 0.0
2722
- },
2723
- "temperature": 1.6817928305074288,
2724
- "calibrated_clean": {
2725
- "n": 243,
2726
- "nll": 0.4460227427836622,
2727
- "acc": 0.8683127572016461,
2728
- "ece": 0.04490914804947624,
2729
- "brier": 0.21677989484802473,
2730
- "mean_conf": 0.9016047263613403,
2731
- "confident_error_rate": 0.024691358024691357,
2732
- "coverage_at_0_9": 0.6831275720164609,
2733
- "accuracy_at_0_9": 0.963855421686747,
2734
- "coverage_at_5pct_error": 0.7489711934156379,
2735
- "coverage_at_1pct_error": 0.4156378600823045,
2736
- "aurc": 0.03366327644063936,
2737
- "error_rate_at_0_9": 0.03614457831325301,
2738
- "confidence_bias": 0.033291969159694124,
2739
- "top_bins": {
2740
- "0.9": {
2741
- "n": 166,
2742
- "errors": 6,
2743
- "error_rate": 0.03614457831325301
2744
- },
2745
- "0.95": {
2746
- "n": 141,
2747
- "errors": 2,
2748
- "error_rate": 0.014184397163120567
2749
- },
2750
- "0.99": {
2751
- "n": 81,
2752
- "errors": 0,
2753
- "error_rate": 0.0
2754
- }
2755
- },
2756
- "selective": {
2757
- "0.5": {
2758
- "coverage": 0.5020576131687243,
2759
- "accuracy": 0.9836065573770492,
2760
- "confidence_cutoff": 0.9739727890189261
2761
- },
2762
- "0.8": {
2763
- "coverage": 0.8024691358024691,
2764
- "accuracy": 0.9282051282051282,
2765
- "confidence_cutoff": 0.8229542441861438
2766
- }
2767
- }
2768
- },
2769
- "metric_policy": {
2770
- "version": 2,
2771
- "selective_ties": "whole_confidence_groups",
2772
- "coverage_at_error": "in-sample maximum over confidence thresholds; not a deployed error guarantee",
2773
- "aurc": "right-step integral over whole confidence groups",
2774
- "confident_error_rate": "high-confidence errors divided by all questions",
2775
- "error_rate_at_0_9": "errors divided by questions accepted at p_max >= 0.9",
2776
- "nll": "exact from logits when recorded; otherwise from floored probabilities",
2777
- "nll_floor": 1e-09,
2778
- "renormalize_returned_probabilities": true,
2779
- "raw_sums_outside_1e_5": 0,
2780
- "returned_zeros": 0
2781
- },
2782
- "coverage": {
2783
- "requested_records": 185,
2784
- "requested_questions": 284,
2785
- "evaluated_records": 185,
2786
- "evaluated_questions": 284,
2787
- "rejected_records": 0,
2788
- "truncated_records": 0
2789
- },
2790
- "latency_ms": {
2791
- "median": 49.463423999895895,
2792
- "p95": 189.5706840000457
2793
- },
2794
- "calibration": {
2795
- "inference_temperature": 1.0,
2796
- "additional_temperature": 1.6817928305074288,
2797
- "logits_recorded": true
2798
- },
2799
- "suite_sha256": "adde94227c959c3bfd62ee2ca27d8a73d5b0fff96bf0500c627645e5a7d72f37"
2800
- },
2801
  "provenance": {
2802
  "config": {
2803
- "epochs": 1,
2804
  "seed": 4,
2805
- "lr": 2.5e-06,
2806
  "lora": 16,
2807
  "accum": 2,
2808
  "batch": 4,
@@ -2823,7 +249,7 @@
2823
  "focal_gamma": 0.0,
2824
  "base": "Qwen/Qwen3.5-4B-Base",
2825
  "base_revision": "1001bb4d826a52d1f399e183466143f4da7b741b",
2826
- "init_from": "macmacmacmac/Sev-4B@a1824aefba305fda86e3503b895a7d9b3871f79a",
2827
  "lora_targets": "all",
2828
  "head_dim": 256,
2829
  "weights_dtype": "fp32",
@@ -2832,8 +258,8 @@
2832
  "option_isolation": 0,
2833
  "special_embeddings": 0
2834
  },
2835
- "config_sha256": "01ce0c9f84977e5c32d40dbd5c012634208c7770615b9e7d9d6f5950c042706f",
2836
- "suite_sha256": "3b9a80dfd212b6932b66ea520c35ab1a3b19e132a81af1b6b7b306d776a70a02",
2837
  "source_hashes": {
2838
  "kev/__init__.py": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
2839
  "kev/anchors.py": "6963eafdb276db5a6c94d939eae675c761555553b8d0f963ffd803448beaabbb",
@@ -2865,62 +291,59 @@
2865
  "pyproject.toml": "39ec40f14c88598dee7ac98fb859a5a44e22337235a3064ebd5e89036ce81e30",
2866
  "uv.lock": "a9922dbb89acdef78299fd2b4a8c3f7f0fa1b2bc08b55595b6926fa785a9c466"
2867
  },
2868
- "git_commit": "740b511bb562cf29430a85457b9e54389a3f21cd",
2869
  "platform": "Linux-4.19.0-gvisor-x86_64-with-glibc2.36",
2870
  "torch": "2.8.0+cu128",
2871
  "device": "cuda",
2872
  "gpu": "NVIDIA H100 80GB HBM3",
2873
  "legacy_checkpoint": false,
2874
  "measured_checkpoint": {
2875
- "requested": "/runs/sev-r2-response-policy-4b-v2/00-trial-0/checkpoint",
2876
- "resolved": "/runs/sev-r2-response-policy-4b-v2/00-trial-0/checkpoint",
2877
- "head_sha256": "506dcd3587847917a288d95757c497c15a2aef8e2d4bc6523b04ad991c961a2b",
2878
- "adapter_sha256": "fa1916b34fe1eeb10d15c36927c225df43115407d7c97b4f879bba3102c3ed21",
2879
  "inference_temperature": 1.0
2880
  }
2881
  },
2882
  "mechanism_checks": {
2883
- "n": 24,
2884
- "packed_max_delta": 1.923413947224617e-06,
2885
- "sibling_max_delta": 9.5367431640625e-07,
2886
  "tolerance": 0.001,
2887
- "passed": true
2888
  },
2889
  "gates": {
2890
- "passed": true,
2891
  "checks": {
2892
  "complete_coverage": true,
2893
- "isolation_and_packing": true,
2894
- "transfer_complete": true,
2895
- "heldout_pairs_at_least_70pct": true,
2896
- "transfer_confident_errors_below_10pct": true
2897
  },
2898
  "policy": "Research screening only: 0.001 isolation; 5pp task regression; 70% heldout pair correctness; <=10% confident errors; no automatic release."
2899
  },
2900
  "calibration_fit": {
2901
- "temperature": 1.6817928305074288,
2902
  "aggregation": "micro",
2903
  "objective": "raw-logit NLL",
2904
  "split": "calibration",
2905
- "rows_sha256": "026b68705b7ac76bc4361bc2fda1cbfe74265b436a807ea68bca77f87d3cf9da",
2906
- "suite_sha256": "3b9a80dfd212b6932b66ea520c35ab1a3b19e132a81af1b6b7b306d776a70a02",
2907
- "n": 2674
2908
  },
2909
- "wall_seconds": 1285.6242614819998,
2910
  "promotable": false,
2911
  "test_evaluated": false,
2912
  "training_resources": {
2913
- "wall_seconds": 887.5156674385071,
2914
- "records_seen": 6577,
2915
- "requested_records": 6577,
2916
  "truncated_records": 0,
2917
  "rejected_records": 0,
2918
- "optimizer_steps": 823,
2919
- "forward_tokens": 1505379,
2920
- "peak_device_bytes": 22225622528,
2921
  "device": "cuda",
2922
  "dtype": "bf16",
2923
  "batch": 4,
2924
- "peak_rss_bytes": 31556661248
2925
  }
2926
  }
 
1
  {
2
+ "objective": -0.8137446728855184,
3
  "paired_flip": null,
4
  "unknowable": null,
5
  "clean": {
6
+ "n": 175,
7
+ "nll": 0.8137446728855184,
8
+ "acc": 0.9028571428571428,
9
+ "ece": 0.09141608822222473,
10
+ "brier": 0.17599452427593348,
11
+ "mean_conf": 0.9942732310793674,
12
+ "confident_error_rate": 0.08,
13
+ "coverage_at_0_9": 0.9828571428571429,
14
+ "accuracy_at_0_9": 0.9186046511627907,
15
+ "coverage_at_5pct_error": 0.9142857142857143,
16
+ "coverage_at_1pct_error": 0.011428571428571429,
17
+ "aurc": 0.06143423339655433,
18
+ "error_rate_at_0_9": 0.08139534883720931,
19
+ "confidence_bias": 0.09141608822222458,
20
  "top_bins": {
21
  "0.9": {
22
+ "n": 172,
23
+ "errors": 14,
24
+ "error_rate": 0.08139534883720931
25
  },
26
  "0.95": {
27
+ "n": 172,
28
+ "errors": 14,
29
+ "error_rate": 0.08139534883720931
30
  },
31
  "0.99": {
32
+ "n": 168,
33
+ "errors": 11,
34
+ "error_rate": 0.06547619047619048
35
  }
36
  },
37
  "selective": {
38
  "0.5": {
39
+ "coverage": 0.5028571428571429,
40
+ "accuracy": 0.9431818181818182,
41
+ "confidence_cutoff": 0.9999983862993777
42
  },
43
  "0.8": {
44
+ "coverage": 0.8,
45
+ "accuracy": 0.9571428571428572,
46
+ "confidence_cutoff": 0.9999883543164546
47
  }
48
  }
49
  },
50
  "tasks": {
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
51
  "sev_swarmtraces_response_policy": {
52
  "n": 175,
53
+ "nll": 0.8137446728855184,
54
+ "acc": 0.9028571428571428,
55
+ "ece": 0.09141608822222473,
56
+ "brier": 0.17599452427593348,
57
+ "mean_conf": 0.9942732310793674,
58
+ "confident_error_rate": 0.08,
59
+ "coverage_at_0_9": 0.9828571428571429,
60
+ "accuracy_at_0_9": 0.9186046511627907,
61
+ "coverage_at_5pct_error": 0.9142857142857143,
62
+ "coverage_at_1pct_error": 0.011428571428571429,
63
+ "aurc": 0.06143423339655433,
64
+ "error_rate_at_0_9": 0.08139534883720931,
65
+ "confidence_bias": 0.09141608822222458,
66
  "top_bins": {
67
  "0.9": {
68
+ "n": 172,
69
+ "errors": 14,
70
+ "error_rate": 0.08139534883720931
71
  },
72
  "0.95": {
73
+ "n": 172,
74
  "errors": 14,
75
+ "error_rate": 0.08139534883720931
76
  },
77
  "0.99": {
78
+ "n": 168,
79
+ "errors": 11,
80
+ "error_rate": 0.06547619047619048
81
  }
82
  },
83
  "selective": {
84
  "0.5": {
85
  "coverage": 0.5028571428571429,
86
+ "accuracy": 0.9431818181818182,
87
+ "confidence_cutoff": 0.9999983862993777
88
  },
89
  "0.8": {
90
  "coverage": 0.8,
91
+ "accuracy": 0.9571428571428572,
92
+ "confidence_cutoff": 0.9999883543164546
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
93
  }
94
  }
95
  }
96
  },
97
  "variants": {
98
  "clean": {
99
+ "n": 175,
100
+ "nll": 0.8137446728855184,
101
+ "acc": 0.9028571428571428,
102
+ "ece": 0.09141608822222473,
103
+ "brier": 0.17599452427593348,
104
+ "mean_conf": 0.9942732310793674,
105
+ "confident_error_rate": 0.08,
106
+ "coverage_at_0_9": 0.9828571428571429,
107
+ "accuracy_at_0_9": 0.9186046511627907,
108
+ "coverage_at_5pct_error": 0.9142857142857143,
109
+ "coverage_at_1pct_error": 0.011428571428571429,
110
+ "aurc": 0.06143423339655433,
111
+ "error_rate_at_0_9": 0.08139534883720931,
112
+ "confidence_bias": 0.09141608822222458,
113
  "top_bins": {
114
  "0.9": {
115
+ "n": 172,
116
+ "errors": 14,
117
+ "error_rate": 0.08139534883720931
118
  },
119
  "0.95": {
120
+ "n": 172,
121
+ "errors": 14,
122
+ "error_rate": 0.08139534883720931
123
  },
124
  "0.99": {
125
+ "n": 168,
126
+ "errors": 11,
127
+ "error_rate": 0.06547619047619048
128
  }
129
  },
130
  "selective": {
131
  "0.5": {
132
+ "coverage": 0.5028571428571429,
133
+ "accuracy": 0.9431818181818182,
134
+ "confidence_cutoff": 0.9999983862993777
135
  },
136
  "0.8": {
137
+ "coverage": 0.8,
138
+ "accuracy": 0.9571428571428572,
139
+ "confidence_cutoff": 0.9999883543164546
140
  }
141
  }
142
  }
 
147
  "mean_max_delta": null,
148
  "flip_rate": null
149
  },
150
+ "temperature": 4.0,
151
  "calibrated_clean": {
152
+ "n": 175,
153
+ "nll": 0.26348233904029333,
154
+ "acc": 0.9028571428571428,
155
+ "ece": 0.04076488602051544,
156
+ "brier": 0.1426360448938433,
157
+ "mean_conf": 0.9436220288776582,
158
+ "confident_error_rate": 0.045714285714285714,
159
+ "coverage_at_0_9": 0.8971428571428571,
160
+ "accuracy_at_0_9": 0.9490445859872612,
161
+ "coverage_at_5pct_error": 0.9142857142857143,
162
+ "coverage_at_1pct_error": 0.011428571428571429,
163
+ "aurc": 0.06143423339655433,
164
+ "error_rate_at_0_9": 0.050955414012738856,
165
+ "confidence_bias": 0.04076488602051542,
166
  "top_bins": {
167
  "0.9": {
168
+ "n": 157,
169
+ "errors": 8,
170
+ "error_rate": 0.050955414012738856
171
  },
172
  "0.95": {
173
+ "n": 131,
174
+ "errors": 5,
175
+ "error_rate": 0.03816793893129771
176
  },
177
  "0.99": {
178
+ "n": 0,
179
  "errors": 0,
180
+ "error_rate": null
181
  }
182
  },
183
  "selective": {
184
  "0.5": {
185
+ "coverage": 0.5028571428571429,
186
+ "accuracy": 0.9431818181818182,
187
+ "confidence_cutoff": 0.9655851033024723
188
  },
189
  "0.8": {
190
+ "coverage": 0.8,
191
+ "accuracy": 0.9571428571428572,
192
+ "confidence_cutoff": 0.9448068121971666
193
  }
194
  }
195
  },
 
207
  "returned_zeros": 0
208
  },
209
  "coverage": {
210
+ "requested_records": 35,
211
+ "requested_questions": 175,
212
+ "evaluated_records": 35,
213
+ "evaluated_questions": 175,
214
  "rejected_records": 0,
215
  "truncated_records": 0
216
  },
217
  "latency_ms": {
218
+ "median": 240.12951999998222,
219
+ "p95": 301.12752519999043
220
  },
221
  "calibration": {
222
  "inference_temperature": 1.0,
223
+ "additional_temperature": 4.0,
224
  "logits_recorded": true
225
  },
226
+ "transfer": null,
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
227
  "provenance": {
228
  "config": {
229
+ "epochs": 5,
230
  "seed": 4,
231
+ "lr": 1e-05,
232
  "lora": 16,
233
  "accum": 2,
234
  "batch": 4,
 
249
  "focal_gamma": 0.0,
250
  "base": "Qwen/Qwen3.5-4B-Base",
251
  "base_revision": "1001bb4d826a52d1f399e183466143f4da7b741b",
252
+ "init_from": "macmacmacmac/Sev-4B@v0.3.0-response-policy-research",
253
  "lora_targets": "all",
254
  "head_dim": 256,
255
  "weights_dtype": "fp32",
 
258
  "option_isolation": 0,
259
  "special_embeddings": 0
260
  },
261
+ "config_sha256": "31d8cf5af214ae4ed0012d1e4fea9fd73e207209cfce1f1112107e4aad65cc9f",
262
+ "suite_sha256": "40d3a8b3bd416a3754ecc30e9791a9ebb14a26c8f2a4422057f769a2dbce4dee",
263
  "source_hashes": {
264
  "kev/__init__.py": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
265
  "kev/anchors.py": "6963eafdb276db5a6c94d939eae675c761555553b8d0f963ffd803448beaabbb",
 
291
  "pyproject.toml": "39ec40f14c88598dee7ac98fb859a5a44e22337235a3064ebd5e89036ce81e30",
292
  "uv.lock": "a9922dbb89acdef78299fd2b4a8c3f7f0fa1b2bc08b55595b6926fa785a9c466"
293
  },
294
+ "git_commit": "fcd6756a329f42a515b4b48f83fb8bde9776232f",
295
  "platform": "Linux-4.19.0-gvisor-x86_64-with-glibc2.36",
296
  "torch": "2.8.0+cu128",
297
  "device": "cuda",
298
  "gpu": "NVIDIA H100 80GB HBM3",
299
  "legacy_checkpoint": false,
300
  "measured_checkpoint": {
301
+ "requested": "/runs/sev-html-contrast-4b-v1/00-trial-0/checkpoint",
302
+ "resolved": "/runs/sev-html-contrast-4b-v1/00-trial-0/checkpoint",
303
+ "head_sha256": "2724bebcc54d4846194130594b98d3e3b9c632935c0f1d5077e46e23c48b2dc8",
304
+ "adapter_sha256": "b58b3963175e6c40bd022db5893ebefb6a6f1e1bd1ec11efc6692a089e1dcf51",
305
  "inference_temperature": 1.0
306
  }
307
  },
308
  "mechanism_checks": {
309
+ "n": 40,
310
+ "packed_max_delta": 0.0015369057655334473,
311
+ "sibling_max_delta": 0.0007038116455078125,
312
  "tolerance": 0.001,
313
+ "passed": false
314
  },
315
  "gates": {
316
+ "passed": false,
317
  "checks": {
318
  "complete_coverage": true,
319
+ "isolation_and_packing": false
 
 
 
320
  },
321
  "policy": "Research screening only: 0.001 isolation; 5pp task regression; 70% heldout pair correctness; <=10% confident errors; no automatic release."
322
  },
323
  "calibration_fit": {
324
+ "temperature": 4.0,
325
  "aggregation": "micro",
326
  "objective": "raw-logit NLL",
327
  "split": "calibration",
328
+ "rows_sha256": "725dc3c48e167e782825490bebc3b3f07b5933660fdca7ac65ad37f13481eb6e",
329
+ "suite_sha256": "40d3a8b3bd416a3754ecc30e9791a9ebb14a26c8f2a4422057f769a2dbce4dee",
330
+ "n": 140
331
  },
332
+ "wall_seconds": 366.39384709,
333
  "promotable": false,
334
  "test_evaluated": false,
335
  "training_resources": {
336
+ "wall_seconds": 232.53715324401855,
337
+ "records_seen": 615,
338
+ "requested_records": 615,
339
  "truncated_records": 0,
340
  "rejected_records": 0,
341
+ "optimizer_steps": 80,
342
+ "forward_tokens": 209595,
343
+ "peak_device_bytes": 20358582784,
344
  "device": "cuda",
345
  "dtype": "bf16",
346
  "batch": 4,
347
+ "peak_rss_bytes": 31542779904
348
  }
349
  }
train.log CHANGED
@@ -1,88 +1,14 @@
1
  Warning: You are sending unauthenticated requests to the HF Hub. Please set a HF_TOKEN to enable higher rate limits and faster downloads.
2
- delta: warm start from /__modal/volumes/vo-RociVuGycNGtzvUcdOqecz/hub/models--macmacmacmac--Sev-4B/snapshots/a1824aefba305fda86e3503b895a7d9b3871f79a: 496 adapter tensors and the pointer head loaded
3
  device=cuda trainable params=33.8M
4
- 6577 training requests (holdout=[]), questions by type {'choice': 8590, 'noul': 518, 'score': 348}
5
  [transformers] `causal_conv1d_fn` is falling back to its reference PyTorch implementation because `causal_conv1d` is not installed. This is correct but much slower; install `causal_conv1d` for the optimized kernel.
6
- ep0 step 10/823 loss 0.397 kl 0.000 anchor 0.000 1.763s/rec
7
- ep0 step 20/823 loss 0.612 kl 0.000 anchor 0.000 1.071s/rec
8
- ep0 step 30/823 loss 0.332 kl 0.000 anchor 0.000 0.761s/rec
9
- ep0 step 40/823 loss 0.428 kl 0.000 anchor 0.000 0.606s/rec
10
- ep0 step 50/823 loss 0.487 kl 0.000 anchor 0.000 0.508s/rec
11
- ep0 step 60/823 loss 0.391 kl 0.000 anchor 0.000 0.442s/rec
12
- ep0 step 70/823 loss 0.304 kl 0.000 anchor 0.000 0.394s/rec
13
- ep0 step 80/823 loss 0.458 kl 0.000 anchor 0.000 0.358s/rec
14
- ep0 step 90/823 loss 0.344 kl 0.000 anchor 0.000 0.330s/rec
15
- ep0 step 100/823 loss 0.675 kl 0.000 anchor 0.000 0.308s/rec
16
- ep0 step 110/823 loss 0.543 kl 0.000 anchor 0.000 0.290s/rec
17
- ep0 step 120/823 loss 0.438 kl 0.000 anchor 0.000 0.274s/rec
18
- ep0 step 130/823 loss 0.481 kl 0.000 anchor 0.000 0.262s/rec
19
- ep0 step 140/823 loss 0.608 kl 0.000 anchor 0.000 0.250s/rec
20
- ep0 step 150/823 loss 0.313 kl 0.000 anchor 0.000 0.241s/rec
21
- ep0 step 160/823 loss 0.419 kl 0.000 anchor 0.000 0.232s/rec
22
- ep0 step 170/823 loss 0.520 kl 0.000 anchor 0.000 0.225s/rec
23
- ep0 step 180/823 loss 0.635 kl 0.000 anchor 0.000 0.218s/rec
24
- ep0 step 190/823 loss 0.429 kl 0.000 anchor 0.000 0.212s/rec
25
- ep0 step 200/823 loss 0.445 kl 0.000 anchor 0.000 0.207s/rec
26
- ep0 step 210/823 loss 0.404 kl 0.000 anchor 0.000 0.203s/rec
27
- ep0 step 220/823 loss 0.472 kl 0.000 anchor 0.000 0.199s/rec
28
- ep0 step 230/823 loss 0.464 kl 0.000 anchor 0.000 0.195s/rec
29
- ep0 step 240/823 loss 0.333 kl 0.000 anchor 0.000 0.192s/rec
30
- ep0 step 250/823 loss 0.507 kl 0.000 anchor 0.000 0.188s/rec
31
- ep0 step 260/823 loss 0.411 kl 0.000 anchor 0.000 0.185s/rec
32
- ep0 step 270/823 loss 0.573 kl 0.000 anchor 0.000 0.182s/rec
33
- ep0 step 280/823 loss 0.492 kl 0.000 anchor 0.000 0.179s/rec
34
- ep0 step 290/823 loss 0.430 kl 0.000 anchor 0.000 0.177s/rec
35
- ep0 step 300/823 loss 0.388 kl 0.000 anchor 0.000 0.182s/rec
36
- ep0 step 310/823 loss 0.426 kl 0.000 anchor 0.000 0.180s/rec
37
- ep0 step 320/823 loss 0.361 kl 0.000 anchor 0.000 0.177s/rec
38
- ep0 step 330/823 loss 0.430 kl 0.000 anchor 0.000 0.175s/rec
39
- ep0 step 340/823 loss 0.510 kl 0.000 anchor 0.000 0.173s/rec
40
- ep0 step 350/823 loss 0.274 kl 0.000 anchor 0.000 0.171s/rec
41
- ep0 step 360/823 loss 0.412 kl 0.000 anchor 0.000 0.170s/rec
42
- ep0 step 370/823 loss 0.459 kl 0.000 anchor 0.000 0.168s/rec
43
- ep0 step 380/823 loss 0.348 kl 0.000 anchor 0.000 0.166s/rec
44
- ep0 step 390/823 loss 0.367 kl 0.000 anchor 0.000 0.165s/rec
45
- ep0 step 400/823 loss 0.523 kl 0.000 anchor 0.000 0.163s/rec
46
- ep0 step 410/823 loss 0.358 kl 0.000 anchor 0.000 0.162s/rec
47
- ep0 step 420/823 loss 0.448 kl 0.000 anchor 0.000 0.161s/rec
48
- ep0 step 430/823 loss 0.474 kl 0.000 anchor 0.000 0.159s/rec
49
- ep0 step 440/823 loss 0.429 kl 0.000 anchor 0.000 0.158s/rec
50
- ep0 step 450/823 loss 0.375 kl 0.000 anchor 0.000 0.157s/rec
51
- ep0 step 460/823 loss 0.413 kl 0.000 anchor 0.000 0.156s/rec
52
- ep0 step 470/823 loss 0.444 kl 0.000 anchor 0.000 0.154s/rec
53
- ep0 step 480/823 loss 0.409 kl 0.000 anchor 0.000 0.154s/rec
54
- ep0 step 490/823 loss 0.272 kl 0.000 anchor 0.000 0.153s/rec
55
- ep0 step 500/823 loss 0.660 kl 0.000 anchor 0.000 0.152s/rec
56
- ep0 step 510/823 loss 0.596 kl 0.000 anchor 0.000 0.151s/rec
57
- ep0 step 520/823 loss 0.453 kl 0.000 anchor 0.000 0.150s/rec
58
- ep0 step 530/823 loss 0.486 kl 0.000 anchor 0.000 0.150s/rec
59
- ep0 step 540/823 loss 0.548 kl 0.000 anchor 0.000 0.149s/rec
60
- ep0 step 550/823 loss 0.412 kl 0.000 anchor 0.000 0.148s/rec
61
- ep0 step 560/823 loss 0.419 kl 0.000 anchor 0.000 0.147s/rec
62
- ep0 step 570/823 loss 0.424 kl 0.000 anchor 0.000 0.147s/rec
63
- ep0 step 580/823 loss 0.378 kl 0.000 anchor 0.000 0.146s/rec
64
- ep0 step 590/823 loss 0.443 kl 0.000 anchor 0.000 0.145s/rec
65
- ep0 step 600/823 loss 0.624 kl 0.000 anchor 0.000 0.145s/rec
66
- ep0 step 610/823 loss 0.536 kl 0.000 anchor 0.000 0.144s/rec
67
- ep0 step 620/823 loss 0.460 kl 0.000 anchor 0.000 0.143s/rec
68
- ep0 step 630/823 loss 0.491 kl 0.000 anchor 0.000 0.143s/rec
69
- ep0 step 640/823 loss 0.472 kl 0.000 anchor 0.000 0.142s/rec
70
- ep0 step 650/823 loss 0.327 kl 0.000 anchor 0.000 0.142s/rec
71
- ep0 step 660/823 loss 0.451 kl 0.000 anchor 0.000 0.141s/rec
72
- ep0 step 670/823 loss 0.380 kl 0.000 anchor 0.000 0.141s/rec
73
- ep0 step 680/823 loss 0.432 kl 0.000 anchor 0.000 0.140s/rec
74
- ep0 step 690/823 loss 0.261 kl 0.000 anchor 0.000 0.140s/rec
75
- ep0 step 700/823 loss 0.610 kl 0.000 anchor 0.000 0.139s/rec
76
- ep0 step 710/823 loss 0.463 kl 0.000 anchor 0.000 0.139s/rec
77
- ep0 step 720/823 loss 0.530 kl 0.000 anchor 0.000 0.138s/rec
78
- ep0 step 730/823 loss 0.495 kl 0.000 anchor 0.000 0.138s/rec
79
- ep0 step 740/823 loss 0.609 kl 0.000 anchor 0.000 0.137s/rec
80
- ep0 step 750/823 loss 0.398 kl 0.000 anchor 0.000 0.137s/rec
81
- ep0 step 760/823 loss 0.536 kl 0.000 anchor 0.000 0.137s/rec
82
- ep0 step 770/823 loss 0.606 kl 0.000 anchor 0.000 0.136s/rec
83
- ep0 step 780/823 loss 0.474 kl 0.000 anchor 0.000 0.136s/rec
84
- ep0 step 790/823 loss 0.423 kl 0.000 anchor 0.000 0.135s/rec
85
- ep0 step 800/823 loss 0.567 kl 0.000 anchor 0.000 0.135s/rec
86
- ep0 step 810/823 loss 0.632 kl 0.000 anchor 0.000 0.135s/rec
87
- ep0 step 820/823 loss 0.528 kl 0.000 anchor 0.000 0.134s/rec
88
- saved /runs/sev-r2-response-policy-4b-v2/00-trial-0/checkpoint
 
1
  Warning: You are sending unauthenticated requests to the HF Hub. Please set a HF_TOKEN to enable higher rate limits and faster downloads.
2
+ delta: warm start from /__modal/volumes/vo-RociVuGycNGtzvUcdOqecz/hub/models--macmacmacmac--Sev-4B/snapshots/da0a131c2636b06c16bbca7d8ae84f0dc1e90b7e: 496 adapter tensors and the pointer head loaded
3
  device=cuda trainable params=33.8M
4
+ 123 training requests (holdout=[]), questions by type {'choice': 188}
5
  [transformers] `causal_conv1d_fn` is falling back to its reference PyTorch implementation because `causal_conv1d` is not installed. This is correct but much slower; install `causal_conv1d` for the optimized kernel.
6
+ ep0 step 10/80 loss 0.521 kl 0.000 anchor 0.000 1.937s/rec
7
+ ep1 step 20/80 loss 0.157 kl 0.000 anchor 0.000 1.087s/rec
8
+ ep1 step 30/80 loss 0.001 kl 0.000 anchor 0.000 0.763s/rec
9
+ ep2 step 40/80 loss 0.000 kl 0.000 anchor 0.000 0.611s/rec
10
+ ep3 step 50/80 loss 0.000 kl 0.000 anchor 0.000 0.518s/rec
11
+ ep3 step 60/80 loss 0.000 kl 0.000 anchor 0.000 0.453s/rec
12
+ ep4 step 70/80 loss 0.000 kl 0.000 anchor 0.000 0.409s/rec
13
+ ep4 step 80/80 loss 0.000 kl 0.000 anchor 0.000 0.377s/rec
14
+ saved /runs/sev-html-contrast-4b-v1/00-trial-0/checkpoint
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
training_config.json CHANGED
@@ -2,8 +2,8 @@
2
  "args": {
3
  "base": "Qwen/Qwen3.5-4B-Base",
4
  "n_per_source": 1000,
5
- "epochs": 1,
6
- "lr": 2.5e-06,
7
  "head_lr": 0.0,
8
  "weight_decay": 0.01,
9
  "lora": 16,
@@ -15,7 +15,7 @@
15
  "label_smoothing": 0.0,
16
  "brier_w": 0.0,
17
  "focal_gamma": 0.0,
18
- "suite": "/root/evals/sev/response-policy-curriculum-v2",
19
  "train_sources": "",
20
  "device": "cuda",
21
  "batch": 4,
@@ -36,19 +36,19 @@
36
  "anchor": "",
37
  "anchor_w": 0.0,
38
  "anchor_sources": "",
39
- "out": "/runs/sev-r2-response-policy-4b-v2/00-trial-0/checkpoint",
40
  "data": "",
41
  "replay": 0,
42
- "init_from": "macmacmacmac/Sev-4B@a1824aefba305fda86e3503b895a7d9b3871f79a",
43
  "seed": 4
44
  },
45
- "suite_sha256": "3b9a80dfd212b6932b66ea520c35ab1a3b19e132a81af1b6b7b306d776a70a02",
46
  "base_revision": "1001bb4d826a52d1f399e183466143f4da7b741b",
47
  "init_source": {
48
- "init_from": "macmacmacmac/Sev-4B@a1824aefba305fda86e3503b895a7d9b3871f79a",
49
- "resolved": "/__modal/volumes/vo-RociVuGycNGtzvUcdOqecz/hub/models--macmacmacmac--Sev-4B/snapshots/a1824aefba305fda86e3503b895a7d9b3871f79a",
50
- "adapter_sha256": "b5afd73584d1098cccb0a1bd42afc40e2de3c3bc3074a5173a642b6d4b880420",
51
- "head_sha256": "1d50c37b4db60c22ae8e6bb24f5055056701bb16e332964350d76df57dc82af0",
52
  "adapter_tensors": 496
53
  },
54
  "ordinal_objective": "ranked_probability_score",
 
2
  "args": {
3
  "base": "Qwen/Qwen3.5-4B-Base",
4
  "n_per_source": 1000,
5
+ "epochs": 5,
6
+ "lr": 1e-05,
7
  "head_lr": 0.0,
8
  "weight_decay": 0.01,
9
  "lora": 16,
 
15
  "label_smoothing": 0.0,
16
  "brier_w": 0.0,
17
  "focal_gamma": 0.0,
18
+ "suite": "/root/evals/sev/response-policy-html-contrast-v1",
19
  "train_sources": "",
20
  "device": "cuda",
21
  "batch": 4,
 
36
  "anchor": "",
37
  "anchor_w": 0.0,
38
  "anchor_sources": "",
39
+ "out": "/runs/sev-html-contrast-4b-v1/00-trial-0/checkpoint",
40
  "data": "",
41
  "replay": 0,
42
+ "init_from": "macmacmacmac/Sev-4B@v0.3.0-response-policy-research",
43
  "seed": 4
44
  },
45
+ "suite_sha256": "40d3a8b3bd416a3754ecc30e9791a9ebb14a26c8f2a4422057f769a2dbce4dee",
46
  "base_revision": "1001bb4d826a52d1f399e183466143f4da7b741b",
47
  "init_source": {
48
+ "init_from": "macmacmacmac/Sev-4B@v0.3.0-response-policy-research",
49
+ "resolved": "/__modal/volumes/vo-RociVuGycNGtzvUcdOqecz/hub/models--macmacmacmac--Sev-4B/snapshots/da0a131c2636b06c16bbca7d8ae84f0dc1e90b7e",
50
+ "adapter_sha256": "fa1916b34fe1eeb10d15c36927c225df43115407d7c97b4f879bba3102c3ed21",
51
+ "head_sha256": "12112262e1822fa4ab2981650bd10ceac46929bb29fbbdd843a39ec9de7dbe82",
52
  "adapter_tensors": 496
53
  },
54
  "ordinal_objective": "ranked_probability_score",
training_metrics.json CHANGED
@@ -1,14 +1,14 @@
1
  {
2
- "wall_seconds": 887.5156674385071,
3
- "records_seen": 6577,
4
- "requested_records": 6577,
5
  "truncated_records": 0,
6
  "rejected_records": 0,
7
- "optimizer_steps": 823,
8
- "forward_tokens": 1505379,
9
- "peak_device_bytes": 22225622528,
10
  "device": "cuda",
11
  "dtype": "bf16",
12
  "batch": 4,
13
- "peak_rss_bytes": 31556661248
14
  }
 
1
  {
2
+ "wall_seconds": 232.53715324401855,
3
+ "records_seen": 615,
4
+ "requested_records": 615,
5
  "truncated_records": 0,
6
  "rejected_records": 0,
7
+ "optimizer_steps": 80,
8
+ "forward_tokens": 209595,
9
+ "peak_device_bytes": 20358582784,
10
  "device": "cuda",
11
  "dtype": "bf16",
12
  "batch": 4,
13
+ "peak_rss_bytes": 31542779904
14
  }