Qev-0.8B

Qev-0.8B is the smallest Qev decision model. Give it context, a question and candidate answers; receive a choice and a probability for every option. It supports Choice, Noul (yes/no) and Score (ordered ratings) through the same API as Qev-2B, Qev-4B and Qev-9B.

Source and installation · Training method · 中文训练说明

This model starts directly from Qwen3.5-0.8B-Base, with fresh rank-64 LoRA, a decision head and a final-layer interaction gate. It learns Qev-9B v0.3.0 option probabilities for two epochs. There is no supervised warm-start, hard-label mixture or representation-response continuation.

Property	Qev-0.8B v0.1.0
Base	Qwen3.5-0.8B-Base, revision dc7cdfe2ee4154fa7e30f5b51ca41bfa40174e68
Adaptation	LoRA rank 64, alpha 128
Decision head	256 dimensions, four attention heads, two Transformer layers
Candidate interaction	Full sibling interaction in the final full-attention layer
Teacher	Qev-9B v0.3.0
Training	Teacher-probability cross entropy; 44,576 single-question inputs
Schedule	Two epochs, 2,786 steps, two GPUs, global batch 32, seed 17
Temperatures	Teacher 1.563437713227029; student 1
Input limits	4,096-token state and complete path; question 512, candidate 256
Computation	BF16 backbone, FP32 decision head
Download	Approximately 192 MiB; the Qwen base downloads separately

Use the model

Install the Qev source package with Python 3.12 and a hardware-compatible PyTorch 2.8.0 build:

from qev import Qev

model = Qev.from_pretrained("AustinFu/Qev-0.8B", revision="v0.1.0", device="cuda")
answers = model.predict({
    "state": "I was charged twice for the same order.",
    "questions": {
        "department": {
            "type": "choice",
            "instructions": "Which team should handle this?",
            "criteria": {"billing": "Charges and refunds", "shipping": "Delivery problems"},
        }
    },
})
print(answers["department"]["choice"])
print(answers["department"]["probabilities"])

For JSONL requests:

python -m qev.predict --checkpoint AustinFu/Qev-0.8B@v0.1.0 \
  --input examples/requests.jsonl --out runs/0.8b-predictions.jsonl \
  --device cuda --weights-dtype checkpoint

The loader restores the adapter, decision head, interaction gate and tokenizer, and downloads the pinned Qwen base. No teacher is needed for inference. The package includes training/fine-tuning configurations, evaluation results and license files. Optimizer state and base weights are excluded.

Recorded evaluation

Benchmark	Correct / total	Accuracy (%)
JevBench public	169 / 231	73.16
Decision development · clean	1029 / 1264	81.41
Transfer development · clean	431 / 656	65.70
MMLU-Pro	247 / 1000	24.70
SemIf · handwritten	109 / 144	75.69
scienthoon	596 / 873	68.27
WANLI	169 / 256	66.02

Results use BF16 backbone computation, an FP32 decision head, temperature 1 and full causal reference execution. All 4,844 questions in the seven complete suites were answered. Development and SemIf rows use the matched subsets in the Qev evaluation guide; evaluation.json preserves the full-suite counts. GSM8K, ChessBench and BPoMP have not been measured for this checkpoint.

Both tested distillation seeds reached 169/231 on JevBench; this release selects seed 17. Distillation improved probability quality but did not beat supervised training on every accuracy metric. The Qev releases have different data and objectives, so their results do not isolate model size alone. NanoJev-0.6B is also included in the source repository as a differently trained small-model reference.

Training and availability

The 44,576-input pool and teacher targets are shared with Qev-4B: general decisions, science and reasoning, Principle judgments, controlled boundaries, web actions and additional rule/reasoning tasks. Each record contains one question. Training uses a single pool with no late-stage repetition or online option edits.

The full training input pool and cached teacher outputs are not distributed. Qev-train remains a separate release of 2,442 synthetic examples with hard labels and synthesis documentation. The source repository provides a runnable training example using the original teacher and user data.

Use configs/qev-0.8b-finetune.json with python -m qev.train --init-checkpoint AustinFu/Qev-0.8B@v0.1.0 for supervised adaptation. Validate performance and probabilities on your own task; calibration and production reliability have not been established.

Qev adaptation weights use Apache-2.0. The Qwen base, tokenizer and source datasets retain their respective terms. License and attribution.