OR-decision-4B

OR-decision-4B is a rank-16 LoRA adapter for Qwen/Qwen3.5-4B, trained to answer typed decisions: yes/no (noul), multiple choice (choice), and ordinal ratings (score). This repository contains the trained adapter, tokenizer, and a small inference helper. Loading it also downloads the original Qwen base model.

The model represents each supplied option as a single choice-identifier token. It scores those identifiers at the end of a fixed prompt and normalizes their logits over the supplied options. This is a specialized vocabulary-readout classifier; ordinary chat generation is not the evaluated inference procedure.

Usage

Install the packages in requirements.txt, then run:

import sys
import torch
from huggingface_hub import snapshot_download
from transformers import AutoModelForImageTextToText, AutoTokenizer
from peft import PeftModel

folder = snapshot_download("alphaXiv/OR-decision-4B")
sys.path.insert(0, folder)
from decision import decision_probabilities

tokenizer = AutoTokenizer.from_pretrained(folder, trust_remote_code=False)
base = AutoModelForImageTextToText.from_pretrained(
    "Qwen/Qwen3.5-4B",
    dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=False,
)
model = PeftModel.from_pretrained(base, folder).eval()

probabilities = decision_probabilities(
    model,
    tokenizer,
    context="The customer needs a copy of an invoice.",
    instructions="Which department should handle this request?",
    options=[
        ("billing", "Invoices and payments"),
        ("technical_support", "Software and device problems"),
    ],
    kind="choice",
)
print(probabilities)
print("Decision:", max(probabilities, key=probabilities.get))

For yes/no questions use kind="noul" and options labeled false and true. For ordinal ratings use kind="score" and ordered, numeric option labels. The helper returns the probability distribution; applications can calculate its expected ordinal value or apply a decision threshold. Keep the supplied option descriptions and full context. The helper rejects inputs exceeding the evaluated 4,096-token window and supports up to 62 distinct option labels.

Training and checkpoint selection

• Backbone: Qwen/Qwen3.5-4B.
• LoRA rank: 16; learning rate: 0.00001; global batch size: 128.
• Corpus: 117,186 training examples after a content-hash split of 123,339 source examples. The mixture includes classification, reasoning, adequacy judgments, and synthetic examples, with frozen Qwen teacher distributions where available.
• Objective: gold-label loss, teacher KL weight 1.0, and frozen-base KL weight 0.3. Ordinal gold targets include adjacent-label smoothing.
• Trained across two epochs. The released checkpoint is update 925 of the second epoch, selected by minimum gold-label negative log likelihood on a fixed 1,024-example validation set. It is not the checkpoint with the highest validation accuracy.
• Checkpoint export uses Spindle/Miles' PEFT export. Optimizer state and the base model weights are not included in this adapter repository.

Evaluation

The released checkpoint was evaluated using Spindle's native vocabulary readout. The exported Transformers/PEFT runtime has not been separately rebenchmarked; runtime, precision, and batching can produce numerical differences.

Evaluation	Result
Fixed validation set	754 / 1,024, 73.63%
Validation gold-label NLL	0.6549
JevBench, primary 4,096-token window	193 / 231, 83.55%
JevBench, secondary 3,072-token window	194 / 231, 83.98%
Base Qwen3.5-4B, primary JevBench window	176 / 231, 76.19%

JevBench used the public tasks and original scoring implementation pinned at 1bcc55eb6c8cffde2306b3db03ede39b61c6152a. Raw and temperature-calibrated correctness counts were identical. Temperature fitting used a separate 4,000-example calibration subset, with a separate 4,000-example calibration test. No training updates occurred during evaluation. The helper above returns raw probabilities, without this fitted temperature.

Limitations

The validation set was reused to select among checkpoints, so its score is not an independent test estimate. The public benchmark has also been reused across experiments. Results are from one training seed and 231 benchmark tasks; small accuracy differences should not be treated as robust improvements. Exact-input overlap checks do not establish absence of semantic contamination. Confidence values are not guaranteed to be calibrated in a new application.

This adapter was trained and evaluated on text decisions. The base model's multimodal capabilities were not evaluated as part of this release.

License

Apache-2.0. The base model is distributed separately by Qwen under its own Apache-2.0 license. See LICENSE and the Qwen3.5-4B model card.