| # Jev Rules-7 (evaluation only) |
|
|
| Seven rule and policy decision tasks (admission policy, rental policy, rules and scenarios (ShARC), agent rules, authority |
| rules, fraud points, patent phrases) plus a missing-fact test, in the same style as the policy tasks that other System One |
| models report. Comparable tasks appear in the jev/jevos project README; the sets used there are not published, so these |
| are rebuilt. Facts in the state, the policy RULE IN THE QUESTION. Code: `rules7.py`. |
|
|
| ## Skills (3,440 items, seed 0) |
|
|
| | skill | items | type | source | licence | |
| |---|---|---|---|---| |
| | `admission` | 400 | noul, 200 yes / 200 no | generated: policy over 2-4 conditions (GPA, exam, letters, experience, age, deadline, portfolio, fee, degree; and / or / unless), thresholds and values near the boundary, distractor fields | CC0-1.0 | |
| | `rental` | 400 | choice (5 outcomes, 80 each) | generated: refuse / approve / surcharge / deposit / both from age, licence years, DUI, vehicle class, payment method | CC0-1.0 | |
| | `agent_rules` | 400 | noul, 200 / 200 | generated: three rules (spend, email, delete, deploy, read data) in the question, one requested action in the state | CC0-1.0 | |
| | `authority` | 400 | choice (5 roles, 80 each) | generated: approval ladders by amount, travel, licences, contracts, hires | CC0-1.0 | |
| | `fraud_points` | 400 | score (4 levels, 100 each) | generated: six point rules (6 of 10, random thresholds / points) summed, level cut-offs in the question | CC0-1.0 | |
| | `sharc` | 400 | noul, 200 / 200 | ShARC dev, yes / no rows only: `tasksource/sharc@fa4605de8281290c69af9dea6118b37d072c2b59` (validation) | CC-BY-SA-3.0 | |
| | `patent_phrases` | 400 | score (0, .25, .5, .75, 1; 80 each) | US Patent Phrase-to-Phrase Matching, held-out split with expert scores: `tasksource/patent-phrase-similarity@ff1868796e44ac9ec02e870b447fde6c5bf2d6e4` (test) | CC-BY-4.0 | |
| | `missing_facts` | 640 (160 pairs per task) | admission (noul) and fraud_points (score) | the same generators; each pair = a complete record and the record with ONE needed fact removed (the rule cannot be decided; chosen by three-valued evaluation) | CC0-1.0 | |
| |
| Split tags: `generated` or `public`. ShARC question wording: "Using only the rule text, the scenario and the conversation |
| so far, do the rules say that the answer to the user's question is yes?" (state: rule text, scenario, conversation, |
| question). The ShARC dev split has only 56 rule texts with yes / no answers, so up to 8 utterances per rule text are |
| kept. Patent state: phrase A, phrase B, CPC class; the five levels are the Kaggle rating scale. |
| |
| ## Sources and exclusions |
| |
| - ShARC and the patent phrases use only eval / test splits at pinned revisions. 36 ShARC dev rows (2 duplicated rule texts, listed in `drop_keys.json`) are excluded from the selection. |
|
|
| ## Metrics |
|
|
| Per skill accuracy / Brier / ECE, `per_split`, and in `extras`: |
| `balanced_accuracy` for the yes / no skills and `missing_facts` -> `auroc/admission`, `auroc/fraud_points`: AUROC of the |
| model's confidence (max probability) separating complete (positive) from missing-fact items, |
| (0.5 = cannot tell). `missing_facts` accuracy is on the complete items only (the missing ones have no gold). |
|
|
| ## Run |
|
|
| `python rules7.py --out DIR` needs network for the two public sets (pinned revisions), a minute; it writes `rules7_v1_items.json.gz`. |
|
|
