aenea-pinta-1.1-mini

AENEA Pinta-1.1 Mini Beta

50M-parameter semantic router for edge dispatch. Pinta-1.1 Mini Beta is the lightweight counterpart to AENEA Pinta-1.1, trading ~3% accuracy for a 4.5× parameter reduction and sub-15 ms CPU inference on commodity hardware. Designed for always-on classification in agent swarms, CLI sidecars, and edge dispatch — where latency matters more than peak accuracy.

Beta notice. This is a Beta release. The routing head carries a measurable class prior from SFT imbalance (~+12 logit spike on <|reserved_27|> Architecture) and requires calibration at inference time. Apply the supplied bias file with scale 0.60 before softmax. See Calibration below.

---

What is this?

Pinta-1.1 Mini is not a generative language model. It is a semantic router that classifies incoming prompts into one of nine dispatch domains and emits a routing token that downstream systems can use to select the appropriate expert model or tool:

The router is trained to predict which token a full generative model would emit next, but does so in a single forward pass without generating any text.

---

Architecture

---

Benchmarks

Evaluated on the Pinta Gold benchmark (1,020 prompts across 8 scored domains + adversarial immunity suite):

The Mini trades ~3% strict accuracy for 4.5× faster inference and 4.6× smaller footprint. For edge deployment where latency budgets are tight, this is a worthwhile trade.

---

Quickstart

Installation

pip install onnxruntime numpy tokenizers onnx

Python wrapper

Download pinta_router.py, tokenizer.json, and eval_prompts_large.chatml.bias.npy from this repository, then:

import numpy as np
from pinta_router import PintaONNXRouter

router = PintaONNXRouter(
    "aenea_pinta_1.1_mini.onnx",
    "tokenizer.json",
    routing_bias=np.load("eval_prompts_large.chatml.bias.npy"),
    bias_scale=0.60,  # calibrated for chatml template
    template="chatml"
)

class_id, confidence, latency_ms = router.route("Reverse a linked list in Python.")
token = router.class_to_token(class_id)

print(f"Route to: {token} (confidence: {confidence:.3f})")
# Output: Route to: <|reserved_24|> (confidence: 0.486)

Confidence gate

The wrapper includes a confidence gate (default threshold 0.35). If the model's max probability falls below threshold, it automatically falls back to <|reserved_29|> (Knowledge) rather than emitting a low-confidence misroute:

class_id, confidence, _ = router.route("What is the capital of France?")
# confidence: 0.188 → below threshold → fallback to Knowledge

This prevents silent misroutes on ambiguous prompts.

---

Repository Layout

aenea-pinta-1.1-mini/
├── pinta_router.py                        # Python router wrapper with calibration
├── tokenizer.json                         # QT-Cittern-1.0 tokenizer (9,216 vocab)
├── eval_prompts_large.chatml.bias.npy     # Calibrated bias vector (scale 0.60)
├── aenea_pinta_1.1_mini.onnx              # ONNX model (fp32, ~210 MB)
├── aenea_pinta_1.1_mini.onnx.data         # External weight data
├── pinta_router.cpp                       # C++ daemon (edge deployment)
├── README.md                              # This file
└── LICENSE                                # Apache 2.0

Integration Note: Unlike Pinta-1.1 (226M), which ships with a full FastAPI-based routing engine (pinta_engine.py), Pinta-1.1 Mini provides only the classification wrapper. This design allows Mini to integrate seamlessly into existing dispatch systems, agent swarms, or custom CLI sidecars. For a complete out-of-the-box routing engine with async dispatch to Ollama/vLLM/OpenAI endpoints, use the full Pinta-1.1 (226M).

---

Calibration

The routing head carries a baked-in class prior from SFT training imbalance. The supplied eval_prompts_large.chatml.bias.npy file contains the estimated prior vector (9 floats, one per class). Apply it with bias_scale=0.60 before softmax:

logits = logits - bias_vector * 0.60
probs = softmax(logits)

Do not skip calibration. Without it, the model will route ~60% of prompts to Architecture (<|reserved_27|>) regardless of input. The scale factor 0.60 was determined by sweeping 0.0→1.0 on a 40-prompt labeled eval set and selecting the optimum accuracy (72.5%) while maintaining acceptable confidence margins.

Re-calibrating on your own data

If you have a labeled prompt set (format: prompt<TAB>label where label is 23-31 or 0-8), you can re-estimate the bias:

python pinta_router.py --model aenea_pinta_1.1_mini.onnx \
    --tokenizer tokenizer.json \
    --calibrate my_prompts.txt \
    --template chatml

This writes my_prompts.chatml.bias.npy. Then sweep to find the optimal scale:

python pinta_router.py --model aenea_pinta_1.1_mini.onnx \
    --tokenizer tokenizer.json \
    --bias my_prompts.chatml.bias.npy \
    --template chatml \
    --sweep-scale \
    --calibrate my_prompts.txt

The script will print accuracy at each scale and recommend the optimum.

---

Intended Use & Ecosystem

Pinta-1.1 Mini is designed for:

• Agent swarm dispatchers — route user queries to specialized agents (code, math, RAG) at the edge
• CLI sidecars — classify shell commands or log lines in real-time with sub-15 ms latency
• Edge devices — run on Raspberry Pi, Jetson, or commodity x86 with CPU-only inference
• Cost-aware routing — send simple prompts to cheap models, complex prompts to expensive ones
• Adversarial defense — detect and redirect prompt injection attempts via the adversarial immunity head

Not intended for:

• Generative text completion (this is a classifier, not a language model)
• High-stakes classification without human oversight (use the confidence gate and fallback)
• Production use without calibration (the prior will cause systematic misroutes)

---

Known Limitations

Beta status

This is a Beta release. The routing accuracy is ~3% below the full Pinta-1.1 (226M) model. Calibration is required.

C++ Daemon: The included pinta_router.cpp provides a baseline CLI runner for edge deployment. Extended C++ daemon bindings and production socket wrappers will follow in the next minor release.

Confidence margins

On the 40-prompt eval set, some prompts (especially factual queries like "What is the capital of France?") fall below the 0.35 confidence threshold even after calibration. These correctly trigger the Knowledge fallback, but may indicate that the 50M model genuinely struggles with borderline cases.

Class imbalance

The SFT training data had uneven class distribution, resulting in the ~+12 logit bias on Architecture. Calibration removes this, but the underlying imbalance means some classes may be under-represented in edge cases.

Template sensitivity

The model was trained with ChatML formatting. Using raw prompts or other templates will degrade accuracy. Always wrap prompts with <|im_start|>user\n{prompt}<|im_end|>.

---

Training Architecture & Datasets

Pinta-1.1 Mini was trained using a two-stage pipeline on consumer-grade hardware, utilizing a STEM-focused corpus with a heavy emphasis on mathematics, code, and technical reasoning.

Training Hardware

• Hardware: NVIDIA RTX 4060 (single GPU for both pre-training and fine-tuning)
• Precision: bfloat16 mixed precision

1. Base Pre-Training (Cartan-50M Checkpoint)

The base Cartan-50M model was pre-trained on a STEM-heavy corpus:

• Mathematics & Reasoning (Dominant): 45+ shards of clean pure/applied math, plus 8 shards of MathOverflow Q&A and 7 shards of Physics reasoning.
• Computer Science & Code: 130+ "sterile" cleaned StackOverflow shards, CodeSearchNet Python, and cleaned source code in C, C++, Python, and Rust. Plus SoftwareEngineering (4 shards), ServerFault (7 shards), SuperUser (15 shards), and AskUbuntu (8 shards).
• General Knowledge: 10 shards of English Wikipedia, 5 shards of general English, and 3 shards of academic papers.
• Specialized Domains: Biology (2 shards), Chemistry (3 shards), Philosophy (1 shard).

2. Fine-Tuning & Contrastive Alignment

• Curated Open Data Mix (50%): Mined imperative subsets targeting instruction execution across Code, System Architecture, Formatting, and Reasoning.
• Deepseek-v4-Flash Synthetic Boundary Data (50%): Hard-negative contrastive pairs synthetically generated to sharpen semantic boundaries between domain classes.

Training Configuration

• Base checkpoint: Cartan-50M with block Householder orthogonal attention
• Training steps: 8,000 steps with gradient accumulation (effective batch size 64)
• Learning rate: 6e-4 with cosine decay to 5% minimum, 500 warmup steps
• Optimizer: AdamW (β1=0.9, β2=0.95, weight_decay=0.1)
• Dataset: english_stem_corpus.txt (500M pretokenized tokens from the STEM corpus above)

Model Card Metadata

• Model type: Semantic router / classifier (20-layer Cartan transformer with causal routing head)
• Base Checkpoint: Cartan-50M
• Inference Size: ~211 MB (ONNX fp32)
• Training Hardware: NVIDIA RTX 4060
• Training procedure: SFT from base Cartan-50M checkpoint, 8000 steps (effective batch size 64 via gradient accumulation), lr=6e-4 with cosine decay to 5% minimum, warmup 500 steps
• Evaluation: Pinta Gold benchmark (strict accuracy, latency)
• Limitations: Requires calibration, Beta accuracy, template-sensitive
• License: Apache 2.0
• Release date: 2026-09-21

---

Links & Contact

• Hugging Face Profile: JamesQuartz
• Full Model: AENEA Pinta-1.1 (226M)
• Company Website: https://aeneaglobal.com/
• Open-Source Models & Tokenizers: quartz.host
• Commercial & Partnership Inquiries: [email protected]

---

Citation

If you use Pinta-1.1 Mini in your research or production systems, please cite:

@misc{pinta-mini-2026,
  title={AENEA Pinta-1.1 Mini: A 50M-Parameter Semantic Router for Edge Dispatch},
  author={Your Name},
  year={2026},
  howpublished={\url{https://huggingface.co/JamesQuartz/aenea-pinta-1.1-mini-beta}},
}

---

License

Apache 2.0. See LICENSE file for details.