blackOrder/Ordivo-v1

Ordivo v1 — Neural Decision Intelligence Model

Ordivo v1 is the first public model release: a PEFT LoRA adapter for single-choice imitation of COPRAS, WASPAS, VIKOR, and TOPSIS decision episodes. It is trained from LiquidAI/d1-3B at immutable revision 051bcc464b01b9f92942b364d9586b0ef5912432.

Training contract

The prompt includes the selected method and parameters, criterion directions and weights, full-precision candidate values, and answer aliases. Completion loss supervises exactly the correct alias and EOS. Restricted option probabilities use d1's upstream alias-group readout. The vision tower is frozen. The adapter targets language attention, convolution, and MLP projections with rank 16, alpha 32, and dropout 0.05. No quantization is used.

Using the published model

The adapter accepts a structured numerical decision problem. It does not accept a free-form business question. Each request supplies an MCDM method, its parameters, weighted benefit/cost criteria, and the value of every alternative on every criterion.

Input fields

Field	Required value
method	COPRAS, WASPAS, VIKOR, or TOPSIS
method_params	{} for COPRAS/TOPSIS, {"lambda": 0.0..1.0} for WASPAS, or {"v": 0.0..1.0} for VIKOR
criteria	Unique name, direction (min or max), and nonnegative weight for each criterion
alternatives	Unique id plus a values mapping containing every criterion name

For inputs closest to the training distribution, use 3–8 alternatives, 2–6 criteria, weights summing to one, and finite strictly positive measurements. Criterion names must match the keys in every alternative's values mapping. The target, solver scores, ranking, provenance, and episode ID are not model inputs.

Sample input

This is published test episode synthetic-v1-42-2-000000 with target fields removed:

decision = {
    "method": "COPRAS",
    "method_params": {},
    "criteria": [
        {
            "name": "criterion_0",
            "direction": "min",
            "weight": 1 / 3,
            "unit": "synthetic",
        },
        {
            "name": "criterion_1",
            "direction": "max",
            "weight": 1 / 3,
            "unit": "synthetic",
        },
        {
            "name": "criterion_2",
            "direction": "min",
            "weight": 1 / 3,
            "unit": "synthetic",
        },
    ],
    "alternatives": [
        {
            "id": "A",
            "values": {
                "criterion_0": 344.958426505142,
                "criterion_1": 174.71343616314272,
                "criterion_2": 4.374510512261219,
            },
        },
        {
            "id": "B",
            "values": {
                "criterion_0": 2074.7917654961116,
                "criterion_1": 40.98290161089295,
                "criterion_2": 24.87462780994877,
            },
        },
        {
            "id": "C",
            "values": {
                "criterion_0": 371.29539511555544,
                "criterion_1": 185.01329033444466,
                "criterion_2": 4.911120661391979,
            },
        },
        {
            "id": "D",
            "values": {
                "criterion_0": 4438.9639476302755,
                "criterion_1": 14.010542557524218,
                "criterion_2": 59.1768121163083,
            },
        },
    ],
}

GPU inference

Use a remote CUDA environment with a compatible PyTorch installation. The release environment used Torch 2.9.1+cu128, Transformers 5.19.0, PEFT 0.21.2, Accelerate 1.15.0, and BF16.

python -m pip install \
  "transformers==5.19.0" \
  "peft==0.21.2" \
  "accelerate==1.15.0" \
  "safetensors==0.8.0" \
  "git+https://github.com/Krishna2709/ordivo.git@v0.1.0"

After defining decision as above:

import torch
from peft import PeftModel
from transformers import AutoModelForImageTextToText, AutoTokenizer

from liquid_mcdm._vendor import d1_prompt
from liquid_mcdm.training.rendering import decision_input, option_probabilities

BASE_MODEL = "LiquidAI/d1-3B"
BASE_REVISION = "051bcc464b01b9f92942b364d9586b0ef5912432"
ADAPTER = "blackOrder/Ordivo-v1"

tokenizer = AutoTokenizer.from_pretrained(
    BASE_MODEL,
    revision=BASE_REVISION,
)
base = AutoModelForImageTextToText.from_pretrained(
    BASE_MODEL,
    revision=BASE_REVISION,
    dtype=torch.bfloat16,
    device_map="auto",
)
model = PeftModel.from_pretrained(base, ADAPTER)
model.eval()

state, question = decision_input(decision)
prompt = d1_prompt.render(
    tokenizer,
    state,
    question,
    bos=tokenizer.bos_token,
)

device = next(model.parameters()).device
inputs = tokenizer(
    prompt,
    add_special_tokens=False,
    return_tensors="pt",
)
inputs = {name: tensor.to(device) for name, tensor in inputs.items()}

with torch.inference_mode():
    logits = model(
        **inputs,
        logits_to_keep=1,
    ).logits[0, -1].float().cpu()

alternative_ids = [item["id"] for item in decision["alternatives"]]
probabilities = option_probabilities(tokenizer, alternative_ids, logits)
result = {
    "winner": max(probabilities, key=probabilities.get),
    "probabilities": probabilities,
}
print(result)

Expected result

The recorded BF16 A40 result for this exact published episode was:

{
    "winner": "A",
    "probabilities": {
        "A": 0.8771265580649198,
        "B": 0.003367434875200803,
        "C": 0.11870617117007114,
        "D": 0.0007998358898083066,
    },
}

Small floating-point differences are possible across GPU and software stacks. The output is one winning alternative and a probability distribution over the provided alternatives; it is not a complete ranking. The probabilities are not calibrated confidence scores. The exact recorded row is available in the release prediction evidence.

Evaluation scope

Run 20261019T005111Z used 512 training episodes, 400 optimizer steps, and an NVIDIA A40 in BF16. Accuracy changed as follows:

Split	Base	Adapted	Episodes
Training subset	32.42%	93.55%	512
Disjoint-family test	28.74%	58.62%	87
Original comparison slice	37.50%	75.00%	16

The full test set contains 100 records; 87 have an admissible unique winner and are eligible for single-choice accuracy. Adapted accuracy by method was 56.0% for COPRAS, 56.0% for TOPSIS, 66.7% for VIKOR, and 60.0% for WASPAS. The release includes configuration, environment freeze, dataset manifest hash, before/after metrics, per-method slices, per-episode predictions, and reload diagnostics. Results are synthetic exact-teacher imitation and do not establish human preference quality, full-ranking ability, procurement suitability, or arithmetic superiority over exact solvers.

The complete training telemetry and release artifacts are in W&B run 4vblgw80.

Known limitations

The release adapter preserved all 16 checked winners after serialization. Repeated inference and two independent fresh reloads were identical. The in-memory-to-reloaded BF16 probability comparison had maximum absolute delta 0.02147, within the preregistered 0.05 decision-equivalence tolerance but not bitwise probability parity. Test calibration error increased from 0.1183 to 0.2394 even as accuracy and Brier score improved, so probabilities should not be interpreted as calibrated confidence. d1-3B and this adapter remain subject to the LFM Open License v1.0; review its terms before use.