Datasets:

LocalLLaMA
/

typed-decisions

@@ -1,55 +0,0 @@
-# Typed Decisions dataset audit
-Gold labels are a **teacher signal** from a ~4B-class endpoint, not ground truth. The benchmark measures how well a small model reproduces a larger one's typed decisions; it does not claim the teacher is correct.
-## agent_trace_observability
-100 cases, 100 distinct states (100.0%), 100% with generated text, 3.00 endpoint attempts per case.
-| Question | Type | Labels | Majority (95% LB) | Entropy | Mean conf | Sample TV | Top factor | MI/H | Flags |
-|---|---|---|---|---|---|---|---|---|---|
-| `outcome` | choice | 4 | success 50% (40%) | 1.16 | 0.46 | 0.27 | violated_constraint | 0.394 | ok |
-| `action` | choice | 4 | human_review 36% (27%) | 1.17 | 0.37 | 0.26 | violated_constraint | 0.214 | ok |
-| `needs_review` | noul | 2 | false 55% (45%) | 0.69 | 0.51 | 0.14 | violated_constraint | 0.278 | ok |
-| `risk` | score | 4 | 0 34% (25%) | 1.35 | 0.54 | 0.24 | irreversible_action | 0.430 | ok |
-| `urgency` | score | 4 | 0 37% (28%) | 1.27 | 0.42 | 0.27 | violated_constraint | 0.347 | ok |
-## customer_service
-100 cases, 100 distinct states (100.0%), 100% with generated text, 3.01 endpoint attempts per case.
-| Question | Type | Labels | Majority (95% LB) | Entropy | Mean conf | Sample TV | Top factor | MI/H | Flags |
-|---|---|---|---|---|---|---|---|---|---|
-| `category` | choice | 5 | billing 38% (29%) | 1.45 | 0.83 | 0.09 | topic | 0.869 | ok |
-| `action` | choice | 4 | escalate_to_human 52% (42%) | 1.17 | 0.50 | 0.32 | topic | 0.363 | ok |
-| `needs_human` | noul | 2 | true 67% (57%) | 0.63 | 0.45 | 0.17 | topic | 0.178 | ok |
-| `urgency` | score | 3 | 2 42% (33%) | 1.02 | 0.57 | 0.23 | tone | 0.236 | ok |
-| `churn_risk` | score | 4 | 2 47% (38%) | 1.15 | 0.59 | 0.19 | topic | 0.352 | ok |
-## invoice_processing
-100 cases, 100 distinct states (100.0%), 100% with generated text, 3.25 endpoint attempts per case.
-| Question | Type | Labels | Majority (95% LB) | Entropy | Mean conf | Sample TV | Top factor | MI/H | Flags |
-|---|---|---|---|---|---|---|---|---|---|
-| `disposition` | choice | 4 | approve 45% (36%) | 1.13 | 0.54 | 0.30 | issue | 0.545 | ok |
-| `matches_order` | noul | 2 | true 54% (44%) | 0.69 | 0.80 | 0.12 | issue | 0.834 | ok |
-| `duplicate` | noul | 2 | false 89% (81%) | 0.35 | 0.83 | 0.09 | issue | 0.681 | ok |
-| `discrepancy_severity` | score | 4 | 0 47% (38%) | 1.06 | 0.63 | 0.23 | issue | 0.615 | ok |
-| `urgency` | score | 3 | 2 51% (41%) | 1.01 | 0.54 | 0.23 | days_until_due | 0.373 | ok |
-## security_incidents
-100 cases, 100 distinct states (100.0%), 96% with generated text, 3.04 endpoint attempts per case.
-| Question | Type | Labels | Majority (95% LB) | Entropy | Mean conf | Sample TV | Top factor | MI/H | Flags |
-|---|---|---|---|---|---|---|---|---|---|
-| `disposition` | choice | 4 | investigate 73% (64%) | 0.85 | 0.42 | 0.21 | rule | 0.368 | ok |
-| `true_positive` | noul | 2 | true 71% (61%) | 0.60 | 0.37 | 0.13 | rule | 0.231 | ok |
-| `credential_compromise` | noul | 2 | true 54% (44%) | 0.69 | 0.31 | 0.14 | rule | 0.175 | ok |
-| `severity` | score | 5 | 3 49% (39%) | 1.22 | 0.49 | 0.22 | rule | 0.177 | ok |
-| `urgency` | score | 4 | 3 49% (39%) | 1.07 | 0.43 | 0.23 | rule | 0.209 | ok |
-**0 flag(s) raised.**
-`MI/H` is the mutual information between the gold label and the most informative sampled latent factor, divided by the label's entropy: the fraction of the label's uncertainty explained by a single known input factor. It is a lower bound on learnability, since a model sees the whole state and not just one factor. A value near zero is the v1 failure -- gold independent of the input.