Decision-4B-v1

Decision-4B-v1 is a 4-billion-parameter structured decision model built from Qwen/Qwen3.5-4B-Base. It was trained with a rank-16 LoRA adapter and a Kev pointer head to score explicitly supplied alternatives rather than generate free-form text.

The model accepts a state and one or more typed questions. It returns calibrated probabilities, a selected choice, a boolean decision, or an expected ordinal score depending on the question type.

This repository contains a Kev checkpoint. It is not a standalone generative chat model. Use the Kev runtime so that both the LoRA adapter and head.pt are loaded correctly.

Model details

Item	Value
Base model	Qwen/Qwen3.5-4B-Base
Checkpoint type	PEFT LoRA adapter plus Kev pointer head
Parameters updated	33.8 million trainable parameters
LoRA rank	16
LoRA targets	All supported projection layers
Decision head dimension	256
Supported question types	choice, noul, score
Training context limit	7,552 state tokens
Primary training language	English
Preliminary evaluated language	Vietnamese
Developed by	Phanhai2804

Intended use

Decision-4B-v1 is intended for applications that can describe a decision as:

1. a factual or policy state;
2. a concise instruction;
3. an explicit set of valid choices, a boolean decision, or an ordered scale.

Suitable uses include intent routing, policy-based decisions, document and clause classification, complaint routing, risk triage, and structured multiple-choice reasoning.

The model is best used as a scoring component inside a larger application. The application should validate the schema, apply domain-specific thresholds, log decisions, and route low-confidence or high-impact cases to human review.

Input and output format

Choice

{
  "state": "A customer reports that the same card payment was charged twice.",
  "questions": {
    "intent": {
      "type": "choice",
      "instructions": "Identify the customer's primary intent.",
      "criteria": {
        "card_payment_duplicate": "The same card payment was charged twice",
        "cash_withdrawal": "The customer wants to withdraw cash",
        "change_pin": "The customer wants to change a PIN"
      }
    }
  }
}

Boolean (noul)

{
  "state": "The policy requires enhanced review for transactions of at least 500,000,000 VND. The transaction is 620,000,000 VND.",
  "questions": {
    "enhanced_review": {
      "type": "noul",
      "instructions": "Does this transaction require enhanced review?"
    }
  }
}

Ordinal score

For score, criteria order defines the scale. The response is an expected position on that scale together with a probability for every level.

{
  "state": "The account completed two successful activities thousands of kilometres apart within five minutes.",
  "questions": {
    "risk": {
      "type": "score",
      "instructions": "Select the immediate system action.",
      "criteria": [
        "allow",
        "monitor",
        "require_otp",
        "block_account"
      ]
    }
  }
}

Do not change the order of score criteria between requests unless the meaning of the ordinal scale is also intended to change.

Running the model

Clone and install the Kev repository:

git clone https://github.com/jaredpalmer/kev.git
cd kev
uv sync --extra serve

Serve a checkpoint downloaded from Hugging Face:

CUDA_VISIBLE_DEVICES=0 \
KEV_DTYPE=bf16 \
uv run --no-sync python -m kev.serve \
  --run Phanhai2804/decision-4b-v1 \
  --device cuda \
  --port 8009

Or serve a local checkpoint directory:

CUDA_VISIBLE_DEVICES=0 \
KEV_DTYPE=bf16 \
uv run --no-sync python -m kev.serve \
  --run runs/decision-data-v1-base-full-20261017-000022 \
  --device cuda \
  --port 8009

Example request:

curl -s http://127.0.0.1:8009/v1/systemone \
  -H "Content-Type: application/json" \
  -d '{
    "state": "The customer says the same card payment was charged twice.",
    "questions": {
      "intent": {
        "type": "choice",
        "instructions": "Identify the primary intent.",
        "criteria": {
          "card_payment_duplicate": null,
          "cash_withdrawal": null,
          "change_pin": null
        }
      }
    }
  }'

The Kev server also exposes GET /v1/models for checkpoint metadata.

Checkpoint contents

File	Purpose
adapter_model.safetensors	LoRA adapter weights
adapter_config.json	PEFT adapter configuration
head.pt	Kev pointer head, temperature, and checkpoint metadata
tokenizer.json	Tokenizer snapshot
tokenizer_config.json	Tokenizer configuration
chat_template.jinja	Input formatting template
training_config.json	Training configuration
training_metrics.json	Recorded training metrics

Loading only the PEFT adapter with PeftModel is insufficient for decision inference because that does not load the Kev pointer head.

Training

Training data

The model was trained on 60,365 structured decision records assembled in three stages:

Component	Records	Purpose
General decision examples	35,785	Multiple-choice and boolean reasoning
Minimal-pair records	17,580	Contrastive choice and none-of-the-above behavior
Complex-domain documents	7,000	Classification and intent-routing tasks
Total	60,365

Training configuration

Hyperparameter	Value
Epochs	2
Learning rate	5e-5
Weight decay	0.01
Batch setting	2
Gradient accumulation	4
LoRA rank	16
Head dimension	256
Compute dtype	BF16
Saved adapter weight dtype	FP32
Gradient checkpointing	Enabled
Maximum state length	7,552 tokens
Length sorting	Enabled
Seed	42

License

The final license is not yet declared. Use of the checkpoint is subject to the base model license and the applicable terms of the training data. Review those terms before redistribution or commercial use.

Contact

Maintainer: Phanhai2804