Datasets:
Shorten dataset card and document original-data model benchmarks
Browse files- README.md +18 -152
- SHA256SUMS +2 -1
- metadata/model-benchmarks.json +109 -0
README.md
CHANGED
|
@@ -33,171 +33,37 @@ configs:
|
|
| 33 |
|
| 34 |
# IMD Oracle Outcomes
|
| 35 |
|
| 36 |
-
|
| 37 |
-
**915,153 panel response records**, collected from the public
|
| 38 |
-
[IMD API](https://api.imd.fun) through **13 October 2026**. Published by Eval Engine;
|
| 39 |
-
this is an independent dataset, not an official IMD release.
|
| 40 |
|
| 41 |
-
|
| 42 |
-
versus disagreed outcomes** from request fields available before execution.
|
| 43 |
-
These labels do not establish factual correctness, question ambiguity, or whether
|
| 44 |
-
the panel reached quorum. The full archive includes all five observed statuses.
|
| 45 |
-
|
| 46 |
-
## October 13 update
|
| 47 |
-
|
| 48 |
-
The full raw archive now contains **7,804 requests** and
|
| 49 |
-
**915,153 panel response records**: **520 new requests**
|
| 50 |
-
relative to October 8. Current statuses: {"attested":7507,"blocked":23,"disagreed":270,"mismatch":3,"refused":1}.
|
| 51 |
-
Unchanged records retain their original fetch timestamps; new or changed records
|
| 52 |
-
are downloaded again. This is an incremental collection, not an atomic snapshot.
|
| 53 |
-
|
| 54 |
-
For the refreshed benchmark, select the dated configuration explicitly:
|
| 55 |
|
| 56 |
```python
|
| 57 |
from datasets import load_dataset
|
| 58 |
updated = load_dataset("evalengine/imd-oracle-outcomes", "outcomes_2026_10_13")
|
| 59 |
```
|
| 60 |
|
| 61 |
-
|
| 62 |
-
| --- | ---: | ---: | ---: |
|
| 63 |
-
| Train | 4,533 | 4,419 | 114 |
|
| 64 |
-
| Validation | 1,349 | 1,343 | 6 |
|
| 65 |
-
| Test | 381 | 311 | 70 |
|
| 66 |
-
|
| 67 |
-
New cutoffs: as-of `2026-10-13T02:58:40.081000+00:00`, validation starts
|
| 68 |
-
`2026-10-04T02:58:40.081000+00:00`, and test starts `2026-10-06T02:58:40.081000+00:00`.
|
| 69 |
-
This version excludes 27 records with other statuses and
|
| 70 |
-
quarantines 1,514 repeated templates.
|
| 71 |
-
|
| 72 |
-
**Versions are separate experiments.** The original `outcomes` configuration and
|
| 73 |
-
its files are unchanged, preserving earlier model comparisons. The dated version
|
| 74 |
-
recomputes temporal splits; old test examples can become new training examples.
|
| 75 |
-
Do not combine versions or reuse earlier scores as results for the new benchmark.
|
| 76 |
-
No models have been evaluated or fine-tuned on this refreshed version.
|
| 77 |
-
|
| 78 |
-
Dated files are under `snapshots/2026-10-13/`; the original audit is preserved at
|
| 79 |
-
`snapshots/2026-10-08/metadata/oracle-audit.json`. The current full-corpus audit is
|
| 80 |
-
`metadata/oracle-audit.json`, and collection provenance is in
|
| 81 |
-
`metadata/refresh-report.json`. All original-response checksums passed, and detail
|
| 82 |
-
IDs and status counts matched a second full API listing and the final API counts.
|
| 83 |
-
Latest audit: 37 normalized templates have both outcomes;
|
| 84 |
-
95 attested records report agreement below quorum.
|
| 85 |
-
|
| 86 |
-
## Where are the answers?
|
| 87 |
-
|
| 88 |
-
The dataset viewer shows questions and outcome labels (`0 = attested`,
|
| 89 |
-
`1 = disagreed`), not oracle answers. Answers are excluded from the benchmark
|
| 90 |
-
inputs so the model cannot see the outcome it is predicting.
|
| 91 |
-
|
| 92 |
-
Actual answers are in [raw/oracle-requests.jsonl.gz](resolve/main/raw/oracle-requests.jsonl.gz),
|
| 93 |
-
under `body.attestation.answer`, `body.agreement.answer`, and
|
| 94 |
-
`body.members[].answer`, where available. These are recorded oracle responses,
|
| 95 |
-
not independently verified ground truth.
|
| 96 |
-
|
| 97 |
-
## Load the benchmark
|
| 98 |
-
|
| 99 |
-
```python
|
| 100 |
-
import json
|
| 101 |
-
from datasets import load_dataset
|
| 102 |
-
|
| 103 |
-
dataset = load_dataset("evalengine/imd-oracle-outcomes", "outcomes")
|
| 104 |
-
example = dataset["train"][0]
|
| 105 |
-
request = json.loads(example["input_json"])
|
| 106 |
-
label = example["label"] # 0 = attested, 1 = disagreed
|
| 107 |
-
```
|
| 108 |
-
|
| 109 |
-
Only `input_json` should be supplied to a prediction model. Other columns contain
|
| 110 |
-
labels or provenance. Request text is untrusted data, not executable instructions.
|
| 111 |
-
|
| 112 |
-
## Contents
|
| 113 |
-
|
| 114 |
-
| Path | Contents |
|
| 115 |
-
| --- | --- |
|
| 116 |
-
| `data/*.jsonl` | Viewer-compatible benchmark splits; request objects serialized losslessly as `input_json` strings |
|
| 117 |
-
| `original/benchmark/*.jsonl` | Unchanged original benchmark files, including quarantine and rejection logs |
|
| 118 |
-
| `raw/oracle-requests.jsonl.gz` | Full 7,804-request archive, including panel responses and provenance envelopes |
|
| 119 |
-
| `metadata/dataset-card.json` | Exact temporal cutoffs, deduplication method, split counts, and canonical hashes |
|
| 120 |
-
| `metadata/oracle-audit.json` | Coverage checks, archive checksum, status distributions, and label audit |
|
| 121 |
-
| `SHA256SUMS` | SHA-256 hashes of uploaded file bytes, excluding this checksum file itself |
|
| 122 |
-
|
| 123 |
-
Each raw archive line is a JSON envelope with the API response in `body`, plus
|
| 124 |
-
its source URL, fetch time, and original response SHA-256. The archive preserves
|
| 125 |
-
parsed source responses; it does not include the separate original HTTP-byte
|
| 126 |
-
cache. The response checksum refers to those original bytes, not a reserialized
|
| 127 |
-
`body` object. This release covers oracle data, not worker submissions or bundles.
|
| 128 |
-
|
| 129 |
-
## Frozen October 8 benchmark: labels and splits
|
| 130 |
-
|
| 131 |
-
| Split | Requests | Attested (0) | Disagreed (1) |
|
| 132 |
-
| --- | ---: | ---: | ---: |
|
| 133 |
-
| Train | 2,652 | 2,563 | 89 |
|
| 134 |
-
| Validation | 1,133 | 1,115 | 18 |
|
| 135 |
-
| Test | 2,156 | 2,137 | 19 |
|
| 136 |
-
|
| 137 |
-
The source snapshot contains 7,089 attested, 172 disagreed, 20 blocked, 2 mismatch,
|
| 138 |
-
and 1 refused request. The benchmark excludes the latter three statuses, retains
|
| 139 |
-
one latest record per normalized question template, and quarantines 1,320 older
|
| 140 |
-
duplicates. Template normalization lowercases text, collapses whitespace, and
|
| 141 |
-
replaces hexadecimal strings and numbers. Templates do not overlap across splits;
|
| 142 |
-
paraphrase leakage remains possible, and normalization can merge questions with
|
| 143 |
-
meaningfully different numeric parameters.
|
| 144 |
-
|
| 145 |
-
Splits use request creation time, with the snapshot cutoff
|
| 146 |
-
`2026-10-08T02:50:29.596Z`: test starts `2026-10-01T02:50:29.596Z`, validation
|
| 147 |
-
starts `2026-09-29T02:50:29.596Z`, and training contains older requests.
|
| 148 |
-
Labels were observed during collection, so this is a retrospective benchmark,
|
| 149 |
-
not a claim that every label was available at its temporal cutoff.
|
| 150 |
-
|
| 151 |
-
## Schema and leakage controls
|
| 152 |
-
|
| 153 |
-
- `id`: source request UUID.
|
| 154 |
-
- `input_json`: JSON string containing an allowlisted request object.
|
| 155 |
-
- `label`: integer, 0 for attested and 1 for disagreed.
|
| 156 |
-
- `observed_status`: source outcome corresponding to the label.
|
| 157 |
-
- `created_at`, `label_observed_at`: request and collection timestamps.
|
| 158 |
-
- `family`: normalized question-template hash.
|
| 159 |
-
- `source_url`, `source_sha256`: source provenance.
|
| 160 |
|
| 161 |
-
The
|
| 162 |
-
`evidence`, `definitions`, `guards`, `panelSize`, `quorum`, `head`,
|
| 163 |
-
`toleranceBps`, `validForSeconds`, and `allowAmbiguous`. Absent fields stay absent.
|
| 164 |
-
Model inputs exclude panel answers, status, attestation, failures, and computed
|
| 165 |
-
results. The API's request fields have not been independently proven immutable
|
| 166 |
-
after submission.
|
| 167 |
|
| 168 |
-
##
|
| 169 |
|
| 170 |
-
|
| 171 |
-
|
| 172 |
-
versus **70 in test** (0.44% versus 18.37%); threshold selection is unstable and
|
| 173 |
-
the temporal class distribution shifts substantially. Accuracy alone is
|
| 174 |
-
misleading; report positive-class recall, precision, average precision, Brier
|
| 175 |
-
score, and the confusion matrix. Select prompts and thresholds using training
|
| 176 |
-
and validation only, and reserve test for the final comparison. Do not oversample
|
| 177 |
-
evaluation splits. Test labels are public in this release, so it is not a hidden
|
| 178 |
-
leaderboard test.
|
| 179 |
|
| 180 |
-
|
| 181 |
-
or policy changes. The October 8 audit found 29 normalized templates with both outcomes,
|
| 182 |
-
and 93 attested records with agreement below quorum. Panel response records are
|
| 183 |
-
not independent questions or human-verified annotations. The archive
|
| 184 |
-
contains post-outcome fields and must not be fed wholesale into an outcome predictor.
|
| 185 |
|
| 186 |
-
|
| 187 |
-
and API status counts: no missing or extra request IDs were found for this
|
| 188 |
-
snapshot. Every cached original response checksum passed before export. Future
|
| 189 |
-
API contents may differ.
|
| 190 |
|
| 191 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 192 |
|
| 193 |
-
|
| 194 |
-
data's redistribution and model-training permissions have not been established.
|
| 195 |
-
This repository does not grant additional rights to the source material. Public
|
| 196 |
-
API access should not be interpreted as an open-data license.
|
| 197 |
|
| 198 |
-
|
| 199 |
|
| 200 |
-
|
| 201 |
-
`2026-10-13T02:58:40.081000+00:00`. When referencing this dataset, cite **Eval Engine,
|
| 202 |
-
IMD Oracle Outcomes, 2026-10-13 update (specify benchmark configuration)**, link this repository and IMD, and
|
| 203 |
-
record the exact repository commit used.
|
|
|
|
| 33 |
|
| 34 |
# IMD Oracle Outcomes
|
| 35 |
|
| 36 |
+
Eval Engine's community dataset of 7,804 IMD oracle requests and 915,153 panel responses, collected from the [IMD API](https://api.imd.fun). Includes questions, recorded answers, outcomes and source provenance.
|
|
|
|
|
|
|
|
|
|
| 37 |
|
| 38 |
+
## Load
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 39 |
|
| 40 |
```python
|
| 41 |
from datasets import load_dataset
|
| 42 |
updated = load_dataset("evalengine/imd-oracle-outcomes", "outcomes_2026_10_13")
|
| 43 |
```
|
| 44 |
|
| 45 |
+
The current benchmark has 4,533 training, 1,349 validation and 381 test examples. Use `input_json` as model input; `label` is `0` for attested and `1` for disagreed. These labels describe protocol outcomes, not factual correctness.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 46 |
|
| 47 |
+
The default `outcomes` configuration retains the original October 8 splits. Keep the two versions separate: their training and evaluation assignments differ.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 48 |
|
| 49 |
+
## Files and answers
|
| 50 |
|
| 51 |
+
- [Full archive](resolve/main/raw/oracle-requests.jsonl.gz): all requests, panel responses and provenance. Available answers appear in `body.attestation.answer`, `body.agreement.answer` and `body.members[].answer`.
|
| 52 |
+
- [Split metadata](blob/main/snapshots/2026-10-13/metadata/dataset-card.json), [audit](blob/main/metadata/oracle-audit.json) and [checksums](blob/main/SHA256SUMS): counts, cutoffs, deduplication and integrity checks.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 53 |
|
| 54 |
+
## Model benchmarks
|
|
|
|
|
|
|
|
|
|
|
|
|
| 55 |
|
| 56 |
+
Task: predict whether an oracle request ends as **attested or disagreed**, using request fields only. All results below use the original dataset's 1,133 validation examples, including 18 disagreements, at a 0.5 threshold.
|
|
|
|
|
|
|
|
|
|
| 57 |
|
| 58 |
+
| Model | Dataset used | Precision | Recall | Average precision |
|
| 59 |
+
| --- | --- | ---: | ---: | ---: |
|
| 60 |
+
| Always attested | Original, Oct 8 | 0.00% | 0.00% | 0.0159 |
|
| 61 |
+
| Qwen3.8-27B FP8 | Original, Oct 8 | 33.33% | 5.56% | 0.0707 |
|
| 62 |
+
| GLM 5.3 Flash | Original, Oct 8 | 8.89% | 22.22% | 0.0587 |
|
| 63 |
+
| GLM 5.3 | Original, Oct 8 | 11.11% | 33.33% | 0.0720 |
|
| 64 |
|
| 65 |
+
Precision and recall measure the disagreement class. These are untuned baselines; reasoning, providers and output settings differ ([results and settings](blob/main/metadata/model-benchmarks.json)). We have not evaluated the new data or the held-out test sets.
|
|
|
|
|
|
|
|
|
|
| 66 |
|
| 67 |
+
Splits use creation time and retain one example per normalized question template. The current validation split has only six disagreements versus 70 in test, so threshold estimates are unstable. Recorded answers have not been independently verified; exclude them from outcome-prediction inputs.
|
| 68 |
|
| 69 |
+
License: unknown. Collected from the publicly accessible IMD API.
|
|
|
|
|
|
|
|
|
SHA256SUMS
CHANGED
|
@@ -1,9 +1,10 @@
|
|
| 1 |
9e75dd981de037ec3769f24f790e126bc5a160b6871f510214e68dc70649aeeb .gitattributes
|
| 2 |
-
|
| 3 |
cbb49b56ea3c650ac26bcb42261601cc70d36c94689f21047aee89cafd51a86f data/test.jsonl
|
| 4 |
776042127eb5332b654c8bf3c73a5b182ea8f86aa77d2211bcecd5dd0a6561e7 data/train.jsonl
|
| 5 |
c430b3d8c59691f64573a8e4ac8db998bb6c9c5c917651814322c720fca8be37 data/validation.jsonl
|
| 6 |
e8906042012cd72a5f933c4aa63a318bd298e5669f170a85f066401a46f6917f metadata/dataset-card.json
|
|
|
|
| 7 |
253447489b29980dbacdb8960b778d33ab10236c50873b93ccd11b53a9b5389e metadata/oracle-audit.json
|
| 8 |
9d02a7e7c586339aa2618e40493ec72c569545cc2d65806687ca399ba2d8b0e3 metadata/refresh-report.json
|
| 9 |
259c0a85dd2fa560ef215a7e879158d60b55d74e6561dd4e200c9d3ede2059e6 original/benchmark/quarantine.jsonl
|
|
|
|
| 1 |
9e75dd981de037ec3769f24f790e126bc5a160b6871f510214e68dc70649aeeb .gitattributes
|
| 2 |
+
5ad8ec9cc799f665b198601b48830c5639936771a76b90deef7fac4e40656784 README.md
|
| 3 |
cbb49b56ea3c650ac26bcb42261601cc70d36c94689f21047aee89cafd51a86f data/test.jsonl
|
| 4 |
776042127eb5332b654c8bf3c73a5b182ea8f86aa77d2211bcecd5dd0a6561e7 data/train.jsonl
|
| 5 |
c430b3d8c59691f64573a8e4ac8db998bb6c9c5c917651814322c720fca8be37 data/validation.jsonl
|
| 6 |
e8906042012cd72a5f933c4aa63a318bd298e5669f170a85f066401a46f6917f metadata/dataset-card.json
|
| 7 |
+
38c780ea5ab231a1356bd81461e055b3e264fca89b8c901a5665cc45034ad17c metadata/model-benchmarks.json
|
| 8 |
253447489b29980dbacdb8960b778d33ab10236c50873b93ccd11b53a9b5389e metadata/oracle-audit.json
|
| 9 |
9d02a7e7c586339aa2618e40493ec72c569545cc2d65806687ca399ba2d8b0e3 metadata/refresh-report.json
|
| 10 |
259c0a85dd2fa560ef215a7e879158d60b55d74e6561dd4e200c9d3ede2059e6 original/benchmark/quarantine.jsonl
|
metadata/model-benchmarks.json
ADDED
|
@@ -0,0 +1,109 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"task": "Predict observed attested versus disagreed outcomes from request fields only",
|
| 3 |
+
"dataset_configuration": "outcomes",
|
| 4 |
+
"dataset_snapshot": "2026-10-08",
|
| 5 |
+
"split": "validation",
|
| 6 |
+
"examples": 1133,
|
| 7 |
+
"positive_class": "disagreed",
|
| 8 |
+
"positive_examples": 18,
|
| 9 |
+
"threshold": 0.5,
|
| 10 |
+
"evaluation_sha256": "3a4c46875f2f18bce7064e8db60460cdbd40ef5f963bab35127477913eeb1dce",
|
| 11 |
+
"train_sha256": "e10e504a46396b09f2bca44add65cda30a54cd10151662d844985a8931241836",
|
| 12 |
+
"models": [
|
| 13 |
+
{
|
| 14 |
+
"model": "Always attested",
|
| 15 |
+
"metrics": {
|
| 16 |
+
"n": 1133,
|
| 17 |
+
"positive_count": 18,
|
| 18 |
+
"disagreement_prevalence": 0.01588702559576346,
|
| 19 |
+
"accuracy": 0.9841129744042365,
|
| 20 |
+
"precision_disagreed": 0.0,
|
| 21 |
+
"recall_disagreed": 0.0,
|
| 22 |
+
"f1_disagreed": 0.0,
|
| 23 |
+
"balanced_accuracy": 0.5,
|
| 24 |
+
"average_precision_disagreed": 0.01588702559576346,
|
| 25 |
+
"brier": 0.01588702559576346,
|
| 26 |
+
"confusion": {
|
| 27 |
+
"tp": 0,
|
| 28 |
+
"tn": 1115,
|
| 29 |
+
"fp": 0,
|
| 30 |
+
"fn": 18
|
| 31 |
+
},
|
| 32 |
+
"threshold": 0.5,
|
| 33 |
+
"warning": null
|
| 34 |
+
}
|
| 35 |
+
},
|
| 36 |
+
{
|
| 37 |
+
"model": "Qwen3.8-27B FP8",
|
| 38 |
+
"metrics": {
|
| 39 |
+
"n": 1133,
|
| 40 |
+
"positive_count": 18,
|
| 41 |
+
"disagreement_prevalence": 0.01588702559576346,
|
| 42 |
+
"accuracy": 0.9832303618711385,
|
| 43 |
+
"precision_disagreed": 0.3333333333333333,
|
| 44 |
+
"recall_disagreed": 0.05555555555555555,
|
| 45 |
+
"f1_disagreed": 0.09523809523809523,
|
| 46 |
+
"balanced_accuracy": 0.5268809167912307,
|
| 47 |
+
"average_precision_disagreed": 0.07070410823444803,
|
| 48 |
+
"brier": 0.01615286849073259,
|
| 49 |
+
"confusion": {
|
| 50 |
+
"tp": 1,
|
| 51 |
+
"tn": 1113,
|
| 52 |
+
"fp": 2,
|
| 53 |
+
"fn": 17
|
| 54 |
+
},
|
| 55 |
+
"threshold": 0.5,
|
| 56 |
+
"warning": null
|
| 57 |
+
}
|
| 58 |
+
},
|
| 59 |
+
{
|
| 60 |
+
"model": "GLM 5.3 Flash",
|
| 61 |
+
"metrics": {
|
| 62 |
+
"n": 1133,
|
| 63 |
+
"positive_count": 18,
|
| 64 |
+
"disagreement_prevalence": 0.01588702559576346,
|
| 65 |
+
"accuracy": 0.9514563106796117,
|
| 66 |
+
"precision_disagreed": 0.08888888888888889,
|
| 67 |
+
"recall_disagreed": 0.2222222222222222,
|
| 68 |
+
"f1_disagreed": 0.12698412698412698,
|
| 69 |
+
"balanced_accuracy": 0.5927254608868959,
|
| 70 |
+
"average_precision_disagreed": 0.058693395682905145,
|
| 71 |
+
"brier": 0.03666557811120812,
|
| 72 |
+
"confusion": {
|
| 73 |
+
"tp": 4,
|
| 74 |
+
"tn": 1074,
|
| 75 |
+
"fp": 41,
|
| 76 |
+
"fn": 14
|
| 77 |
+
},
|
| 78 |
+
"threshold": 0.5,
|
| 79 |
+
"warning": null
|
| 80 |
+
}
|
| 81 |
+
},
|
| 82 |
+
{
|
| 83 |
+
"model": "GLM 5.3",
|
| 84 |
+
"metrics": {
|
| 85 |
+
"n": 1133,
|
| 86 |
+
"positive_count": 18,
|
| 87 |
+
"disagreement_prevalence": 0.01588702559576346,
|
| 88 |
+
"accuracy": 0.9470432480141218,
|
| 89 |
+
"precision_disagreed": 0.1111111111111111,
|
| 90 |
+
"recall_disagreed": 0.3333333333333333,
|
| 91 |
+
"f1_disagreed": 0.16666666666666666,
|
| 92 |
+
"balanced_accuracy": 0.6451420029895366,
|
| 93 |
+
"average_precision_disagreed": 0.0719890311070724,
|
| 94 |
+
"brier": 0.03432753751103268,
|
| 95 |
+
"confusion": {
|
| 96 |
+
"tp": 6,
|
| 97 |
+
"tn": 1067,
|
| 98 |
+
"fp": 48,
|
| 99 |
+
"fn": 12
|
| 100 |
+
},
|
| 101 |
+
"threshold": 0.5,
|
| 102 |
+
"warning": null
|
| 103 |
+
}
|
| 104 |
+
}
|
| 105 |
+
],
|
| 106 |
+
"fine_tuning_performed": false,
|
| 107 |
+
"test_set_evaluated": false,
|
| 108 |
+
"settings_note": "Qwen reasoning disabled, max_tokens 256; GLM models low reasoning, max_tokens 4096. GLM 5.3 uses Sail Research FP8, strict JSON schema and up to two format retries; Flash uses floor routing and JSON object mode. Serving settings differ."
|
| 109 |
+
}
|