Archai JEV Nesso 4B Ita (System One Decision Engine)
Archai JEV Nesso 4B Ita is a highly calibrated, state-of-the-art Italian "System One" decision engine fine-tuned from mii-llm/nesso-4B.
Designed for deterministic, low-latency execution in production edge environments, this model evaluates complex Italian inputs and outputs calibrated single-token decisions (A, B, C, D, TRUE, FALSE) across safety moderation, intent routing, boolean logic, and sentiment analysis.
Key Highlights
- Outstanding Accuracy: Reaches 93.69% Test Accuracy across a multi-task dataset suite in Italian.
- High Probabilistic Calibration: Achieves a 0.0908 Brier Score, ensuring that output logit distributions reliably reflect real confidence.
- Edge Deployment Ready: Quantized to
Q8_0GGUF for execution in C++/Rust (llama.cpp,candle).
Model Details
- Developer: nickprock
- Base Model:
mii-llm/nesso-4B(4 Billion Parameters) - Architecture: LlamaForCausalLM
- Fine-Tuning Method: QLoRA ($r=16, \alpha=32$)
- Quantization Format: GGUF (
Q8_0) - Target Tokens:
A,B,C,D,TRUE,FALSE
Benchmark Results
| Metric | Epoch 1 | Epoch 2 | Final Test Set |
|---|---|---|---|
| Train Loss | 0.2620 | 0.0628 | — |
| Validation Loss | 0.1444 | 0.1903 | — |
| Validation Accuracy | 95.17% | 96.17% | — |
| Test Accuracy | — | — | 93.69% |
| Test Brier Score | — | — | 0.0908 |
Dataset Suite
- Safety & Moderation:
Paul/hatecheck-italian+evalitahf/hatespeech_detection - Intent Routing / Logic:
RiTA-nlp/ai2_arc_ita(ARC-Challenge + ARC-Easy) - Sentiment Classification:
evalitahf/sentiment_analysis - Natural Language Inference:
evalitahf/textual_entailment