Mira-v1

Fast, calibrated decisions. No text generation.

Overview

Mira-v1 is a lightweight decision model developed by SAGEA. It transforms structured state and typed questions into probability distributions in a single forward pass, without autoregressive text generation.

Designed for classification, intent routing, and decision systems, Mira supports choice, score, and noul queries, individually or in batches. It returns calibrated probabilities, confidence estimates, and an abstention signal, enabling downstream systems to make explicit, confidence-aware decisions.

Key features

• Single-pass inference: Produces decision probabilities without generating text.
• Typed queries: Supports choice, score, and noul decision formats.
• Calibrated confidence: Provides probability distributions and abstention signals.
• Batch inference: Evaluates multiple questions in a single forward pass.
• Lightweight deployment: Designed to run efficiently on commodity CPUs.
• API compatibility: Can be served through a TypeSafe-compatible interface.

Quickstart

from transformers import AutoModel, AutoTokenizer
import torch

tok = AutoTokenizer.from_pretrained("sagea-ai/Mira-v1")
enc = AutoModel.from_pretrained("sagea-ai/Mira-v1").eval()

heads = torch.load(
    __import__("huggingface_hub").hf_hub_download(
        "sagea-ai/Mira-v1", "heads.pt"
    ),
    map_location="cpu",
    weights_only=True
)

state = "My payouts have been failing for 3 days."
opts = ["billing", "technical", "sales"]

hs = enc(**tok(
    [state + " [SEP] " + o for o in opts],
    padding=True,
    return_tensors="pt"
)).last_hidden_state[:, 0]

probs = (
    (hs @ heads["scorer"]["weight"].T).squeeze(-1) / 1.05
).softmax(-1)

print(dict(zip(opts, [round(float(p), 3) for p in probs])))

For API deployment, use the TypeSafe-compatible endpoint:

POST /v1/systemone

Evaluation

Mira-v1 is evaluated against open decision models using a shared benchmark harness. Performance measurements are conducted on commodity CPU hardware (Intel Core i5-1335U).

Mira performs competitively on typed decisions and demonstrates substantially improved calibration over the evaluated open baseline. It also performs strongly on intent classification, particularly Banking77 at native full-label resolution.

Topic classification remains a limitation. Mira is designed primarily for decision routing rather than general-purpose topic classification. Emotion classification is more competitive, while knowledge-heavy and shortlist-based comparisons remain areas where Julia and Jev perform better.

Nepali, romanized Nepali, and code-mixed evaluations are also included in Mira's internal reporting.

Limitations

Mira compares candidate answers supplied by the caller. It cannot retrieve missing facts, generate responses, or perform arithmetic.

Severity estimation and urgency separation remain areas for improvement. Beyond 20 candidate options, a two-stage shortlist is recommended. Recall should be evaluated for the intended application.

Training data is private. Evaluation methodology and predictions are publicly available.

Intended use

Mira-v1 is intended for lightweight classification, intent routing, and confidence-aware decision systems. It is not a general-purpose language model and should not be used as the sole basis for consequential decisions without application-specific evaluation.