Datasets:
Tasks:
Text Classification
Formats:
parquet
Languages:
English
Size:
1K - 10K
Tags:
structured-decisions
calibration
probabilistic-classification
system-one
workflow-evaluation
Synthetic
License:
Submission: ProKope-421M-8K (Alan Canto) — 0.7185 acc / 0.385 Brier / 8K context (specialist)
#23
by AlanCantoFTW - opened
Official Submission: ProKope-421M-8K (Alan Canto)
We are officially submitting ProKope-421M-8K, conceived and engineered by Alan Canto (Framework Architecture by Wesley Foreman), for inclusion in the official LocalLLaMA/typed-decisions benchmark ledger.
Operating across a full 8,192-token context window, ProKope-421M-8K is a specialized non-autoregressive System 1 decision engine that processes structured input states in a single forward pass (<25ms execution latency, 0 output tokens).
- Model Repository: ProKope-AI/ProKope-421M-8K
- Creator, Author & Lead Architect: Alan Canto
- Framework Architect: Wesley Foreman
- Architecture: ModernBERT-large (395M) + Geodesic SLERP Multi-Head (26.5M) — 421.3M parameters total
- Context Capacity: 8,192 tokens (full uncompressed document/trace ingestion)
- Mode: Specialist (fitted on
trainsplit) - Precision: Native
bfloat16(206 tensors, 803.57 MB)
Verified Test Set Benchmark Metrics (400 cases / 2,000 decisions):
| Metric | Score / Measurement | Notes |
|---|---|---|
| Overall Accuracy | 71.85% (1,437 / 2,000) | Surpasses Featherless Simple Jev 35B (71.60%) |
| - Noul (Boolean) | 80.50% (483 / 600) | First sub-1B model to cross 80% on boolean verification |
| - Choice (Categorical) | 70.00% (420 / 600) | Calibrated categorical routing |
| - Score (Continuous) | 66.75% (534 / 800) | Continuous confidence rating |
| Decision Brier Score | 0.3847 | Multi-class vector Brier score |
| Score MAE | 0.5205 | Mean Absolute Error |
| ECE | 0.4302 | Expected Calibration Error |
| Inference Latency | 42.94 ms/dec | 23.3 decisions/sec on Sovereign Single-GPU Workstation BF16 |
Domain Breakdown:
- Invoice Processing: 79.80% (399 / 500)
- Agent-Trace Observability: 73.80% (369 / 500)
- Security Incidents: 69.60% (348 / 500)
- Customer Service: 64.20% (321 / 500)
Turnkey Verification Artifacts Published to Repository:
- Official Evaluation Results YAML:
.eval_results/typed-decisions.yaml - Machine-Readable Scorecard:
benchmark_scorecard.json - Empirical Prediction Attestation:
eval_predictions.jsonl(Full 2,000 decision log) - Deterministic Evaluation Script:
eval_prokope.py(1-command re-evaluation)