EpochEcho/decision-4b-preview

Decision 4B Preview

Structured decisions. Explicit criteria. One probability per option.

4B parameters · Text-only · 32K context · Merged BF16 weights

Quick start · Interface · Evaluation · Model details

Decision 4B Preview is a preview release of a decision model fine-tuned from Qwen3.5-4B. Give it a state and questions with explicit answer criteria; it returns structured choices and probabilities without generating explanations.

Use it for intent classification, workflow routing, tool selection, evidence judgments, and other tasks with a defined answer space. Labels are supplied in each request, so the model is not restricted to a fixed set of classes.

Quick start

Requirements: Python 3.13 and a CUDA GPU. The reference stack uses PyTorch 2.14.1, CUDA 13.2, and Transformers 5.17.0. Install a matching CUDA build of PyTorch, then run from this model folder:

pip install -r requirements.txt

from model import DecisionModel

model = DecisionModel("/path/to/model-folder")

response = model.decide({
    "state": "I was charged twice for the same order.",
    "questions": {
        "team": {
            "type": "choice",
            "instructions": "Which team should handle this request?",
            "criteria": {
                "billing": "Payments, invoices, and refunds",
                "technical": "Product bugs and integration issues",
            },
        },
        "payment_issue": {
            "type": "noul",
            "instructions": "Does this request describe a payment problem?",
        },
    },
})

print(response["answers"])

The response contains a selected option and a probability for every criterion in each choice question. A noul answer contains noul, the estimated probability that the proposition is true.

Interface

Field	Description
state	Text or a JSON-compatible value containing the context
questions	A nonempty mapping of question IDs to definitions
type: "choice"	Select from 1–255 named criteria
type: "noul"	Estimate P(true); true/false descriptions are optional
instructions	The decision to make; defaults to the question ID

Each question is evaluated independently, with the complete state. Answers are combined into one response. Independent inputs can be batched with model.decide_many(requests).

The 32,768-token limit includes the state, question, criteria, and formatting for one question. Oversized inputs are refused rather than truncated. All supplied options receive a probability. Temperature is fixed at 1.0, and probabilities are normalized only over valid options.

Evaluation

Public Index 0.3	Score
Overall	57.72
Knowledge & Reasoning	42.12
Language Understanding	65.04
Retrieval & Classification	64.02
Tools & Automation	73.38
Arts & Human Taste	37.83
Successful requests	140178 / 140178

Public Index uses official chance correction and category weights. It is not a simple category average and is not the leaderboard's private Full Score. The complete result files are supplied separately.

Model details

Component	Configuration
Base	Qwen3.5-4B
Training	Supervised fine-tuning on 2 million decision records
Adaptation	LoRA rank 64, alpha 64; merged into the released model weights
Full-attention layers	Bidirectional over each question's complete input
Gated DeltaNet layers	Native causal behavior
Decision head	Linear, 255 answer-code rows × 2,560 hidden dimensions, no bias
Storage	BF16 backbone, FP32 decision head
Runtime	Transformers and PyTorch; no Unsloth or PEFT requirement

Training covers knowledge and reasoning, language understanding, retrieval and classification, tools, preferences, and schema-based decisions. It uses original task supervision, including upstream synthetic and preference annotations. Model selection used a separate development set of 15,935 requests.

The repository contains a complete merged checkpoint, tokenizer, configuration, and linear readout. No base-model download or adapter loading is needed. The included renderer, bidirectional mask, and marker readout are required: ordinary causal text generation does not reproduce this model. File checksums are recorded in manifest.json.

Limits and responsible use

• Text only: no image/video input, free-form generation, or arbitrary score rubrics.
• Probabilities are model estimates, not guaranteed calibrated confidence. Validate thresholds on your own held-out data.
• Multi-question requests repeat the state for each independent question; reported token usage includes that repetition.
• Long contexts need additional GPU working memory beyond the roughly 9 GiB checkpoint footprint.
• Exact and approximate source/group overlap checks were used; they do not prove exhaustive semantic decontamination.
• No private Full Score or official single-request latency qualification is claimed.

For the separately installed official Decision Index kit, the included adapter is engine:ModelEngine.

License

Apache License 2.0. Base-model copyright and license notices are retained. Upstream datasets keep their own terms and are not redistributed in this repository.