Qwen3.5-0.8B Typed Decisions

A task-specific decision model trained locally with Unsloth on Apple Silicon. It combines the Qwen/Qwen3.5-0.8B backbone with a newly initialized Clef-style joint decision head, trained on LocalLLaMA/typed-decisions (the all subset selected in the training UI).

This is an independent fine-tune, not TypeSafe's Jev and not an official Cloudflare Clef checkpoint. The intended interface accepts a state and typed questions and returns probabilities/scores. A normal text-generation pipeline does not execute the decision head.

Intended use

Research and prototyping for the four dataset workflows: agent trace observability, customer service, invoice processing, and security incidents. These are dataset domains, not demonstrated production capabilities. Inputs include state and questions, including criteria; target distributions come from gold.

The training dataset is synthetic and English-language. Chinese-language and multimodal decision quality have not been evaluated for this checkpoint.

Training and saved results

The following values are recorded in the exported unsloth_decision_config.json:

Setting Value
Base Qwen/Qwen3.5-0.8B
Decision head clef (new)
Training method QLoRA
Objective soft_cross_entropy
Epochs 3
Optimizer steps 225
Saved maximum input length 4096
Held-out decisions 2000
Training duration 3970.6 seconds (~66.2 minutes)
Reported peak memory 3.1 GB
Export timestamp 2026-10-18T08:19:08Z

The training UI was configured with learning rate 1e-4, LoRA rank 16, alpha 16, dropout 0, batch size 8, gradient accumulation 2, linear LR schedule, and weight decay 0.01. These settings are transcribed from the final UI screenshot; the exported summary does not independently record all of them.

Hardware: Apple M4 Pro with 48 GiB unified memory. The peak-memory value above is a trainer-reported metric, not a guarantee of total system memory usage.

Trainer-reported metric Before training After training
Held-out accuracy 0.355 0.780
Held-out ECE 0.0289 0.0208

These are results saved by the training application, not independently reproduced benchmarks. The summary does not record the exact evaluation IDs, ECE binning, per-workflow metrics, or whether temperature fitting used the same held-out examples. The model includes calibration settings. Do not treat the reported ECE as an independently established deployment guarantee or compare this fine-tuned result directly with zero-shot leaderboard results. ECE is not the prediction error rate.

Files and format

This export contains merged backbone weights plus a separate decision head, not just a LoRA adapter:

  • model.safetensors-00001-of-00001.safetensors and model.safetensors.index.json: backbone weights and index.
  • joint_head.safetensors and joint_head_config.json: trained decision head.
  • unsloth_decision_config.json: decision runtime, calibration and training summary.
  • config.json, tokenizer files, chat template and processor configuration: supporting model assets.
  • LICENSE: retained Apache-2.0 license.

QLoRA describes the training method; it does not mean the uploaded merged weights are a 4-bit GGUF or a standalone quantized MLX export. Keep all model files together, including the decision head and calibration configuration.

Usage

Use an Unsloth version with Clef decision-model support. The original environment uses unsloth_zoo==2026.10.6, mlx==0.32.3 and mlx-lm==0.31.3. These are recorded environment versions, not a fully pinned dependency lockfile.

In Unsloth Desktop, select the checkpoint in the Decision / System One model selector and use the /v1/systemone interface. Ordinary chat inference does not reproduce the decision-head results.

The following Apple Silicon Python example follows the installed Unsloth MLX API. This example was successfully executed against the local checkpoint on Apple Silicon before publication. The model returned a normalized option distribution and selected investigate (probability approximately 0.8005). This is a functional smoke test, not an additional accuracy benchmark. Use it inside a compatible Unsloth MLX environment. The example uses this repository ID; a local checkpoint directory can also be supplied. Private repositories require Hugging Face authentication.

from unsloth_zoo.mlx.decision import load_decision_model

model = load_decision_model("SnowGemma/Qwen3.5-0.8B-Typed-Decisions")
result = model.answer(
    state="The customer reports that an order is three days overdue and requests help.",
    questions={
        "action": {
            "type": "choice",
            "instructions": "Choose the next support action based on the available evidence.",
            "criteria": {
                "investigate": "Check shipment status and contact the carrier.",
                "refund": "Issue a refund when refund eligibility is established.",
                "close": "Close the case when the issue has been resolved."
            }
        }
    }
)
print(result)

A clean-environment Hub download and inference test remains recommended; the smoke test above used the original local checkpoint and installed runtime. Do not assume support in generic Transformers, Ollama, or hosted inference providers merely because the backbone is Qwen.

Limitations

  • Accuracy is measured against the dataset's teacher-generated targets, not comprehensive human adjudication.
  • Four synthetic workflows do not establish broad decision-making competence.
  • Confidence may be miscalibrated on unfamiliar inputs. Evaluate thresholds on representative deployment data.
  • No independent per-domain, multilingual, vision, adversarial or production evaluation is provided.
  • Software using the outputs remains responsible for business rules and final actions.

License and attribution

The base model is distributed under Apache-2.0; this release retains that license. The training dataset is also labeled Apache-2.0. Runtime software has its own licenses, which are separate from this weight release.

中文简介

这是使用 Unsloth 在 Apple M4 Pro 上训练的 Qwen3.5-0.8B 决策模型,包含新建并训练的 Clef 类型决策头。它接收状态、问题和判断标准,输出选项概率或评分;不是官方 Jev,也不是普通聊天微调模型。训练程序保存的准确率为 0.355 → 0.780,ECE 为 0.0208,这些数值尚未独立复测。使用时必须保留决策头及校准配置,并通过支持该格式的决策运行时加载。

Downloads last month
69
Safetensors
Model size
0.9B params
Tensor type
BF16
·
F32
·
MLX
Hardware compatibility
Log In to add your hardware

Quantized

Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for SnowGemma/Qwen3.5-0.8B-Typed-Decisions

Finetuned
(954)
this model

Dataset used to train SnowGemma/Qwen3.5-0.8B-Typed-Decisions