Matilda-Jev by Maincode

Matilda-Jev is Maincode's one-pass decision model. It scores the options supplied in a choice, noul (yes/no), or ordered score question. It accepts text or JSON state and optional images. This is a decision checkpoint with a 255-option readout, not a text-generation checkpoint.

This configuration edition uses MatildaJevModel, MatildaJevConfig, and MATILDA tokenizer/processor classes. Use the bundled runtime below, or load the custom AutoClasses with trust_remote_code=True. The package includes the backbone weights, decision readout, tokenizer, preprocessing configuration and calibrated temperature.

Validation

On AMD MI355X with Python 3.12, Transformers 5.17.0 and PyTorch 2.14.0:

• 25/25 smoke-test questions passed, including six product-identity questions and an image question.
• All 23 requests matched the original checkpoint's answer probabilities exactly (maximum difference 0).
• Tokenization, image preprocessing, configuration save/reload and architecture-parameter comparisons passed.
• A real forward/backward pass produced finite, nonzero readout and embedding gradients. No optimizer update was applied.

See TEST_REPORT.json. These checks are not a full benchmark rerun or a full continued-training run. Other accelerator backends are untested for this configuration edition.

Results

Evaluated with Decision Index 0.3 (public) over 150,317 requests (currently awaiting official submission).

Area scores are chance-corrected skill multiplied by 100.

Download and serve

Use Python 3.12. Install PyTorch 2.14.0 and torchvision 0.29.0 for your accelerator first. The bf16 weights require approximately 49 GiB before runtime overhead; testing used an AMD MI355X.

python -m pip install huggingface_hub
hf auth login
hf download Maincode/matilda-jev-v1 --local-dir ./matilda-jev-v1
python -m pip install -r ./matilda-jev-v1/requirements-runtime.txt
PYTHONPATH="$(pwd)/matilda-jev-v1/runtime" python -m maincode_jev_serve.server \
  --checkpoint ./matilda-jev-v1 --model-name matilda-jev-v1 \
  --host 127.0.0.1 --port 8000

The API is at http://127.0.0.1:8000/v1/systemone; interactive documentation is at /docs. Set MJ_API_KEY or MJ_API_KEY_FILE to enable authentication. The bundled runtime loads both the backbone and readout.safetensors, and applies the stored temperature.

Example request

Send this JSON to POST /v1/systemone:

{
  "model": "matilda-jev-v1",
  "state": "My debit card was charged twice for the same purchase.",
  "questions": {
    "q": {
      "type": "choice",
      "instructions": "What is the issue?",
      "criteria": {
        "card_delivery": null,
        "duplicate_charge": null,
        "cash_withdrawal": null
      }
    }
  }
}

Recorded response from the validated local package:

{
  "model": "matilda-jev-v1",
  "answers": {
    "q": {
      "type": "choice",
      "probabilities": {
        "card_delivery": 0.0012471709832156106,
        "duplicate_charge": 0.9971489469298866,
        "cash_withdrawal": 0.0016038820868977167
      },
      "choice": "duplicate_charge",
      "confidence": 0.99572342039483
    }
  },
  "usage": {
    "input_tokens": 100,
    "output_tokens": 0
  }
}

AutoClass loading

import torch
from transformers import AutoModel, AutoProcessor

path = "./matilda-jev-v1"
processor = AutoProcessor.from_pretrained(
    path, trust_remote_code=True, local_files_only=True)
backbone = AutoModel.from_pretrained(
    path, trust_remote_code=True, local_files_only=True,
    dtype=torch.bfloat16, attn_implementation="sdpa").to("cuda")

AutoModel returns the backbone hidden states. Use the bundled runtime for calibrated JEV decisions. See USAGE.txt for the decision-model training constructor. Save further training runs into a new output directory.

Licence and attribution

See LICENSE. The custom classes reuse the Transformers implementation.