amaara-decide-vision
A 0.8B vision model that scores options instead of writing text. Given an image, a question and a set of options, it returns a choice with a calibrated probability for every option in a single forward pass. It never generates prose.
Trained by The A-Tech Corporation PTY LTD, Queensland, Australia.
This repository holds the full-precision weights. To run it with llama.cpp, LM Studio or Ollama, use amaara-decide-vision-GGUF, which has F16, Q8_0, Q6_K, Q5_K_M and Q4_K_M builds plus the required vision projector.
Why this shape
Ask a chat model "which segment is largest?" about a chart and you pay for a paragraph of explanation you did not want, and you get a claim you cannot threshold. A decision model returns a distribution over the options and stops, typically in under a second on CPU.
Results
Measured on held-out data only, with a wrong-image ablation on the same rows. The ablation is the point: it shows the model reads the image rather than answering from the question text and option list.
| source | n | correct | mismatched image | delta | ECE | skill* |
|---|---|---|---|---|---|---|
| charts, held-out styles (set A) | 500 | 0.848 | 0.266 | +0.582 | 0.042 | 0.79 |
| charts, held-out styles (set B) | 500 | 0.864 | 0.256 | +0.608 | 0.036 | 0.82 |
UI widgets (rico-widget) |
297 | 0.848 | 0.362 | +0.486 | 0.032 | 0.76 |
map yes/no (mapqa) |
500 | 0.942 | 0.580 | +0.362 | 0.020 | 0.86 |
diagrams (ai2d) |
500 | 0.862 | 0.650 | +0.212 | 0.035 | 0.61 |
geometry (intergps) โ not supported |
323 | 0.390 | 0.347 | +0.043 | 0.059 | 0.07 |
* skill = delta / (1 โ accuracy with a mismatched image) โ the share of the headroom the
image actually buys. Raw accuracy alone overstates the skill: on the diagram set the answer
is already 65% recoverable with a wrong image, so accuracy alone would credit the model
for reading the question.
Two independent held-out chart sets agree to within 0.026. That agreement is the strongest evidence that chart reading is a real transferable capability here rather than a fit to one particular split.
The ablation that mattered
Before this version, the chart family was inverted: 0.325 with the correct chart and 0.450 with another row's chart. A negative delta means the model does worse with the right image than a random one, because its learned bias anti-correlates with what is drawn. The fix was generating 8,000 CC0 business charts with ground truth computed from the underlying numbers โ no value labels in ~75% and non-zero axis starts in ~35%, so the axis must be read. The delta is now +0.58 to +0.61 on styles never trained on.
Files in this repository
| file | size | what it is |
|---|---|---|
model.safetensors |
1.75 GB | merged weights, bf16 |
model-fp8.safetensors |
0.87 GB | FP8 e4m3 backbone; scales in fp8_scales.json |
model-fp8-rest.safetensors |
4 MB | tensors kept out of FP8 (norms, embeddings offset) |
joint_head.safetensors |
48 MB | Clef decision head โ required by every variant |
joint_head_config.json, joint_schema_model.py |
โ | head configuration and schema code |
temperatures.json |
โ | fitted per-bucket calibration scales |
For GGUFs and the vision projector, go to amaara-decide-vision-GGUF.
Usage โ full weights
These are the merged weights with the Clef head intact; load them with the model's own
joint_schema_model.py and Unsloth's FastDecisionModel:
from unsloth import FastDecisionModel
model, processor = FastDecisionModel.from_pretrained(
"theatechcorporation/amaara-decide-vision", load_in_4bit=False
)
answers, report = FastDecisionModel.predict(
model, processor, state="Which region has the highest revenue?",
questions={"q1": {"type": "choice",
"instructions": "Choose the criterion that best answers the question.",
"criteria": {"North": "", "South": "", "East": ""}}},
images=["chart.png"],
)
The head needs max_seq_length >= 4096: the image tokens plus the option schema occupy
about 4k tokens before the state begins.
Calibration temperatures are shipped as temperatures.json and are already baked into the
GGUFs. If you serve these weights directly, apply them โ without them the model is
measurably overconfident.
FP8
model-fp8.safetensors is e4m3 with one scale per tensor (per-row for embeddings) for
vLLM-style FP8 lanes. Norm layers and the 38M Clef head stay bf16 โ the head carries the
calibration and an 8-bit cast would destroy it.
Measured cost of the whole round trip (quantise to e4m3, dequantise, run the same GGUF conversion and evaluation as bf16):
| source | n | bf16 | FP8 round-trip | delta | ECE bf16 โ FP8 |
|---|---|---|---|---|---|
ai2d |
500 | 0.862 | 0.860 | -0.002 | 0.035 โ 0.019 |
mapqa |
500 | 0.942 | 0.944 | +0.002 | 0.020 โ 0.019 |
rico-widget |
296 | 0.848 | 0.851 | +0.003 | 0.032 โ 0.037 |
intergps |
323 | 0.390 | 0.390 | 0.000 | 0.059 โ 0.047 |
FP8 storage costs at most 0.003 accuracy on these tasks. Note that llama.cpp cannot read FP8 safetensors โ use the GGUF repository there.
Training data
15,967 rows over 10,249 unique images:
tasksource/tasksource-jev-typed-decisions, configvision, commercial + unspecified licences only (10,000 non-commercial rows excluded):vision/ai2d(diagrams),vision/mapqa/yesno(maps),vision/rico-widget/element(Android UI),vision/intergps(geometry).- 8,000 synthetic business charts generated for this model (CC0) โ bar, line, pie and tables, with ground truth computed from the numbers.
The train/holdout split is by image hash, not by row. The source dataset has 14,199 rows over only 9,917 unique images โ 43% of rows share an image with another row โ so a row-level split places the same diagram on both sides and measures memorisation. Verified zero image overlap between train and holdout.
| Base model | unsloth/Qwen3.5-0.8B (vision-language, Apache-2.0) |
| Method | Unsloth FastDecisionModel + Clef joint-schema head |
| Fine-tuning | QLoRA rank 16, 4-bit, bf16 compute |
| Trainable | 38.1M of 699M parameters (5.46%) |
| Steps | 2,000 |
| Context | 4,096 tokens |
| Hardware | 1 ร RTX 5060 Ti 16 GB, native Windows 11 |
Limitations โ read before use
- Geometry is not supported.
intergpsis a geometry dataset (triangles, inscribed angles), not business charts. It is listed above only for completeness; its delta is +0.043 and its ECE misses the bar. Do not use this model on geometry. - Charts are validated on synthetic data. The evidence is two independent held-out sets of generated charts, not customer dashboards. Validate on your own imagery before deploying.
- Diagram skill is capped by the question text. On
ai2dthe answer is 65% recoverable with a wrong image, so the image can add at most 35 points and the model captures 61% of that. - Calibration is per option-count bucket. One temperature per bucket corrects scale, not shape. ECE is 0.020โ0.042 on the supported tasks but 0.059 on geometry.
- Single seed, single training length. No variance estimate across seeds.
- No frontier-VLM comparison on the same rows, so there is no efficiency-per-point baseline.
Provenance
- Trained by The A-Tech Corporation PTY LTD, QLD, Australia
- Training data: licence-filtered per row; synthetic charts generated here and released CC0
- Base model: Apache-2.0
Citation
@misc{amaara-decide-vision,
title = {amaara-decide-vision: a calibrated 0.8B vision decision model},
author = {The A-Tech Corporation PTY LTD},
year = {2026}
}
- Downloads last month
- 195
