| --- |
| license: cc-by-nc-sa-4.0 |
| base_model: Qwen/Qwen3.5-4B |
| base_model_relation: adapter |
| tags: |
| - multimodal |
| - vision |
| - decision-model |
| - lora |
| - qwen3_5 |
| --- |
| |
| # Jolt-4B |
|
|
|  |
|
|
| Jolt-4B is a multimodal decision model. Given text or an image and a set of |
| candidate answers, it returns candidate probabilities, a truth estimate, or an |
| ordered score. It does not generate free-form answers. |
|
|
| ## Model details |
|
|
| | Property | Value | |
| | --- | --- | |
| | Base model | [Qwen/Qwen3.5-4B](https://huggingface.co/Qwen/Qwen3.5-4B) | |
| | Base revision | `a7b0d22b993d71000cf2eadfb37222a67cee521e` | |
| | Checkpoint | L2, seed 42, selected at update 3,600 | |
| | Development macro accuracy | 84.39% (checkpoint selection partition) | |
| | Training objective | Direct-composite loss | |
| | Training precision | BF16 base | |
| | Trainable weights | Rank-8 language LoRA and a linear candidate scorer | |
| | Input | Text, or text plus one decoded image | |
| | Limits | 2–128 candidates per question; 4,096-token inference limit | |
|
|
| The reported development score selected the checkpoint within its run. It is not |
| an independent test score. Test data were not used to select this checkpoint. |
|
|
| ### Local evaluation results |
|
|
| | Held-out test lane | Dataset macro | Row weighted | |
| | --- | ---: | ---: | |
| | Text · 18 datasets | 85.11% | 85.91% | |
| | Images · 8 datasets | 83.96% | 96.92% | |
| | All · 26 datasets | 84.76% | 93.47% | |
|
|
| JevBench v1.2.2 public subset: **181/231** (78.35% pooled; 84.50% tier-balanced). The public subset contains 231 tasks; 303 official tasks were unavailable. This is a single checkpoint, seed 42, selected at update 3,600. Unit temperatures were used; no post-training temperature fit was applied. |
|
|
| ### JevBench |
|
|
|  |
|
|
| ### 18 text dataset benchmarks |
|
|
|  |
|
|
| ### 8 image dataset benchmarks |
|
|
|  |
|
|
| You can find a comprehensive model comparison report at: *model_comparison_dashboard.html* |
|
|
| ## Use |
|
|
| This repository contains the compact Jolt adapter and inference source. It does |
| not duplicate Qwen's base weights: the runtime downloads the pinned public Qwen |
| revision on first use and caches it locally. The adapter format is Jolt's custom |
| decision-model format, so load it with the included runtime rather than |
| `AutoModel.from_pretrained`. |
|
|
| The runtime currently requires Python 3.12+, a CUDA-enabled PyTorch setup, and a |
| CUDA GPU. From a clone or downloaded copy of this repository: |
|
|
| ```bash |
| python -m pip install -r requirements-cuda.txt |
| python -m pip install -e . --no-deps |
| ``` |
|
|
| Then: |
|
|
| ```python |
| from jolt.predictor import Predictor |
| |
| model = Predictor.from_checkpoint("mlengineer-ai/Jolt-4B") |
| result = model.predict( |
| {"message": "I was charged twice. Please refund the duplicate."}, |
| { |
| "route": { |
| "type": "choice", |
| "instructions": "Which team should handle this request?", |
| "criteria": ["billing", "technical support", "sales"], |
| }, |
| "refund": { |
| "type": "noul", |
| "instructions": "Is a refund requested?", |
| }, |
| }, |
| ) |
| |
| print(result["answers"]["route"]["choice"]) |
| print(result["answers"]["refund"]["noul"]) |
| ``` |
|
|
| For an image question, pass a decoded PIL image as `state["image"]`. This |
| example uses the `model` loaded above: |
|
|
| ```python |
| from PIL import Image |
| |
| image = Image.open("path/to/image.jpg").convert("RGB") |
| image_result = model.predict( |
| {"image": image, "question": "Which animal is shown?"}, |
| { |
| "animal": { |
| "type": "choice", |
| "instructions": "Identify the animal in the image.", |
| "criteria": ["cat", "dog", "bird", "other"], |
| } |
| }, |
| ) |
| |
| print(image_result["answers"]["animal"]["choice"]) |
| ``` |
|
|
| ## Training and data |
|
|
| Jolt follows the recipe based on the [Dohnuts](https://huggingface.co/PsiACE/Dohnuts-0.1.0-0.8B) project. These checkpoints start from the |
| Qwen3.5 base model listed above; they do not use Dohnuts model weights as their |
| base. Jolt trained rank-8 LoRA weights and a candidate scorer with the base |
| model frozen. Jolt's public-data task suite and split design are adapted from |
| Dohnuts. Its [data reference](https://github.com/PsiACE/dohnuts/blob/main/docs/data-and-evaluation.md) |
| lists source datasets and their known terms. The adapter metadata records |
| fingerprints for Jolt's training, development, calibration, and test partitions. |
| Training data are not included here. |
|
|
| ## Model variations |
|
|
| * https://huggingface.co/mlengineer-ai/Jolt-0.8B |
| * https://huggingface.co/mlengineer-ai/Jolt-2B |
| * https://huggingface.co/mlengineer-ai/Jolt-4B |
| * https://huggingface.co/mlengineer-ai/Jolt-9B |
|
|
| ## Licensing |
|
|
| This release offers the adapter and decision-head weights under CC BY-NC-SA 4.0; |
| see `LICENSE`. For this prepared release, this conservative default follows the |
| original Dohnuts model release. It is not a legal determination that every source |
| dataset's terms necessarily apply to trained weights. The training mixture |
| includes sources with different terms, including |
| [ScienceQA](https://github.com/lupantech/ScienceQA/blob/main/README.md), whose |
| maintainers specify CC BY-NC-SA 4.0 for the dataset. The available records do |
| not establish commercial clearance for these trained weights; review the source |
| dataset terms before use. |
|
|
| The bundled inference source is adapted from Dohnuts and is licensed under |
| Apache-2.0; see `LICENSE-CODE` and `NOTICE-CODE`. The Qwen base model is |
| separately licensed by its publisher under Apache-2.0. Those licenses do not |
| replace or change the terms of training data. |
|
|
