README.md

---
language: en
license: other
license_name: research-use-derived-data
library_name: peft
base_model: jaredpalmer/kev-0.8b
base_model_relation: adapter
pipeline_tag: visual-question-answering
tags:
  - decision-model
  - calibration
  - typesafe
  - kev
  - jev
  - vision
  - qwen3.5
datasets:
  - Jacqkues/kev-vision-decisions-full
  - Jacqkues/sev-ood
  - Jacqkues/sev-generic-bench
---

# Sev-0.8B

**Sev** is a *decision model that reads images*: one image plus a state (text or JSON) and a set of typed questions go in,
a calibrated probability distribution per question comes out, in a single forward pass, with no text generation.
Question types are the TypeSafe `/v1/systemone` ones: `choice` (one of named options), `noul` (yes / no) and `score`
(ordered levels). New questions and options are given at request time; nothing is retrained.

It is [Kev-0.8B](https://huggingface.co/jaredpalmer/kev-0.8b) (a Jev-style text decision model: LoRA + pointer head on
`Qwen/Qwen3.5-0.8B-Base`) plugged into the **full Qwen3.5 vision-language model** and fine-tuned on
[kev-vision-decisions-full](https://huggingface.co/datasets/Jacqkues/kev-vision-decisions-full) (gated: research use, per-source licences). The vision tower is frozen;
the Kev LoRA (r = 16) and pointer head are trained further.

- **Status**: a research project, heavily vibe-coded with Claude Code. Not for production.
- **Document version**: [Sev-0.8B-docindex](https://huggingface.co/Jacqkues/sev-0.8b-docindex) (page indexing: 21 attributes)
- **Code**: `kev-vision` (`VisionDecisionModel`, training, dataset builder, evaluation)
- **Served temperature**: T = 1.32, fitted on the validation split (hard-label questions only)

## Use

```python
from kev.api import SystemOneRequest, to_record
from kev_vision.model import VisionDecisionModel
from PIL import Image

model = VisionDecisionModel("Jacqkues/sev-0.8b", device="cuda").eval()     # Ampere/Ada GPU, MPS or CPU (see limits)
request = {"state": "A document page is attached.",
           "questions": {"doc_type": {"type": "choice", "instructions": "What type of document is this?",
                                      "criteria": {"invoice": "An invoice", "receipt": "A receipt", "letter": "A letter"}},
                         "handwritten": {"type": "noul", "instructions": "Is it mostly handwritten?"}}}
rec, meta = to_record(SystemOneRequest.model_validate(request)); rec["image"] = Image.open("page.jpg")
for m, p in zip(meta, model.probs(rec)):
    print(m["id"], dict(zip(m["keys"], p.tolist())))
```

## How it works

```
[state] <vision_start> image tokens <vision_end> state text  |  [q] instructions [opt] option ... [decide]
Qwen3.5 (vision frozen, LoRA on the language model)          →  pointer head: <decide> · each option  →  softmax / T
```

Every question is its own causal row over the shared state (Qwen3.5 is a Gated DeltaNet hybrid), positions are Qwen's 3D
M-RoPE, and each option is scored at its end token. On text-only input Sev's code reproduces Kev exactly (max |Δp| 5.7e-7).

## Training

- **Start**: `jaredpalmer/kev-0.8b` (adapter + head). A controlled comparison on 20k records picked it over a bare
  Qwen3.5-0.8B-Base with a fresh LoRA of the same rank: better on 14 of 21 test sets, clearly on text (0.827 vs 0.766).
- **Data**: kev-vision-decisions-full, ~126k training records: VQAv2 yes/no + counting, A-OKVQA, ScienceQA, EuroSAT, Oxford
  Pets, FairFace age, AVA aesthetics (soft targets), ChartNet, document types (DocLayNet, CORD receipts, invoices,
  RVL-CDIP), control (Pong, Breakout, Freeway, Space Invaders, Ms. Pac-Man from JAT; LeRobot PushT, upweighted ×2),
  text replay from Kev's sources, and blank / noise images with a uniform target. Yes/no balanced per source, option
  order randomised, images shared between train and held-out splits removed.
- **Recipe**: 1 epoch (136,502 passages), batch 8, AdamW lr 5e-5 OneCycle, bf16 autocast, Kev's augmentation (option
  permutation, none-of-the-above, distractors) redrawn each pass. ~3 h on one A100 80 GB (Modal).
- **Validation**: loss 0.6635, accuracy 0.763, Brier 0.31.

## Results: test splits (200 records each, served temperature)

| test set | before (Kev-0.8B, zero-shot) | **Sev-0.8B** | Brier | ECE |
|---|---|---|---|---|
| `ai2d_ood` | 0.585 | **0.665** | 0.446 | 0.077 |
| `aokvqa` | 0.77 | **0.810** | 0.267 | 0.028 |
| `ava` | 0.23 | **0.640** | 0.706 | 0.397 |
| `breakout` | 0.37 | **0.445** | 0.650 | 0.068 |
| `chartnet` | 0.825 | **0.971** | 0.043 | 0.016 |
| `countbench_ood` | 0.555 | **0.730** | 0.399 | 0.106 |
| `documents` | 0.866 | **0.985** | 0.025 | 0.012 |
| `eurosat` | 0.588 | **0.940** | 0.087 | 0.029 |
| `fairface_age` | 0.295 | **0.595** | 0.544 | 0.059 |
| `freeway` | 0.862 | **0.814** | 0.359 | 0.176 |
| `mspacman` | 0.05 | **0.380** | 0.768 | 0.062 |
| `pets` | 0.815 | **0.975** | 0.036 | 0.025 |
| `pong` | 0.462 | **0.519** | 0.548 | 0.048 |
| `pope_ood` | 0.855 | **0.880** | 0.178 | 0.051 |
| `pusht` | 0.213 | **0.647** | 0.485 | 0.084 |
| `rvl_cdip` | 0.567 | **0.861** | 0.202 | 0.031 |
| `scienceqa` | 0.619 | **0.942** | 0.102 | 0.051 |
| `spaceinvaders` | 0.35 | **0.490** | 0.667 | 0.086 |
| `text` | 0.805 | **0.830** | 0.243 | 0.057 |
| `text_full` | 0.844 | **0.851** | 0.214 | 0.053 |
| `vqav2` | 0.796 | **0.853** | 0.210 | 0.042 |

Blank images (the image carries no answer): mean confidence **0.245** against a uniform 0.238.

## Results: sev-ood (never trained on)

Unseen Atari games, unseen classes, questions the image cannot answer, and corrupted images
([Jacqkues/sev-ood](https://huggingface.co/datasets/Jacqkues/sev-ood), gated, 500 records per set). *Confident errors* = wrong at
p ≥ 0.9; *coverage* = share of decisions that could be automated with at most 5% of them wrong.

| set | accuracy | ECE | confident errors | coverage at 5% error |
|---|---|---|---|---|
| `corrupted_eurosat` | **0.555** | 0.263 | 0.187 | 0.00 |
| `corrupted_pets` | **0.938** | 0.018 | 0.005 | 0.98 |
| `corrupted_vqav2` | **0.767** | 0.039 | 0.025 | 0.21 |
| `ood_beamrider` | **0.350** | 0.081 | 0.000 | 0.00 |
| `ood_boxing` | **0.162** | 0.038 | 0.000 | 0.00 |
| `ood_cars` | **0.872** | 0.054 | 0.000 | 0.75 |
| `ood_flowers` | **0.734** | 0.083 | 0.022 | 0.53 |
| `ood_food101` | **0.886** | 0.010 | 0.012 | 0.81 |
| `ood_qbert` | **0.278** | 0.011 | 0.000 | 0.00 |
| `ood_seaquest` | **0.160** | 0.027 | 0.000 | 0.00 |
| `unknowable` (uniform target) | mean confidence 0.396 vs chance 0.227 | | answered at ≥ 0.9: **1.2%** | |

## Results: generic vision benchmarks (never trained on)

Five public benchmarks, cast as typed questions: MMStar (A-D, vision-indispensable), MMBench-en dev (1,000 sampled),
MME (yes/no), RealWorldQA (A-F or yes/no) and HallusionBench (image split, yes/no). ScienceQA-sourced items were dropped
(Sev trained on ScienceQA). **Contamination control**: every benchmark image was hashed (256-bit average hash) against
every training image; the 305 of 6,001 items within 6 bits of a training image (mostly COCO photos via VQAv2 / A-OKVQA,
and ScienceQA) are removed below. Removing them changes no conclusion: they were slightly easier for all three models,
including the baseline that never saw them.

The baseline is **Qwen3.5-0.8B** (the post-trained VLM of the same family) in zero-shot, scored by its next-token
probability over the option letters (or Yes / No), so the same calibration metrics apply. All models see images capped at
448 x 448 pixels. Accuracy on the 5,696 clean items; gap = Sev minus Qwen, paired bootstrap 95% interval.

| benchmark (chance) | n | Qwen3.5-0.8B | **Sev-0.8B** | gap [95% CI] | ECE Qwen → Sev |
|---|---|---|---|---|---|
| MMStar (0.25) | 1,036 | 0.460 | **0.526** | +6.6 [+3.3, +9.9] | 0.18 → **0.08** |
| MMBench dev (0.26) | 927 | 0.774 | **0.825** | +5.1 [+2.6, +7.6] | **0.02** → 0.06 |
| MME (0.50) | 2,206 | 0.711 | **0.771** | +6.0 [+4.2, +7.9] | 0.12 → **0.01** |
| RealWorldQA (0.39) | 586 | 0.570 | **0.631** | +6.1 [+2.2, +10.2] | 0.15 → **0.05** |
| HallusionBench (0.50) | 941 | 0.644 | 0.629 | −1.5 [−4.8, +1.9] | 0.15 → **0.10** |
| **all** | 5,696 | 0.650 | **0.698** | **+4.7 [+3.5, +6.0]** | 0.12 → **0.02** |

Confident errors (wrong at p ≥ 0.9), all clean items: 4.5% for Qwen, **1.5%** for Sev. HallusionBench (visual illusions,
edited charts) is the exception: no model of this size is reliably better than chance-level guessing there.

## Results: docbench (document pages, never trained on)

DocLayNet test + validation pages, three yes/no questions (table of contents? figure? table?), 676 items, labels from
independent rules. See [Sev-0.8B-docindex](https://huggingface.co/Jacqkues/sev-0.8b-docindex) for the document model.

| model | accuracy | AUROC | ECE |
|---|---|---|---|
| Kev-0.8B | 0.886 | 0.934 | 0.236 |
| Sev-0.8B | 0.831 | 0.924 | **0.044** |
| Sev-0.8B-docindex | **0.928** | **0.976** | **0.018** |

## Limits

- **Control does not transfer**: unseen games stay near chance, and on trained games Sev is still below simply repeating
  the previous action (Pong 0.52 vs ~0.69). It is cautious there (no confident errors), not competent.
- **Degraded satellite images**: on blurred / noised EuroSAT tiles Sev is over-confident (18.7% confident errors).
- **Aesthetics (AVA) is poorly calibrated** (ECE 0.40): a subjective target.
- Zero-shot classification of unseen classes is good (cars 0.87, food 0.89, flowers 0.73) but 1-4 points below the
  20k-record Kev start: more task-specific training erodes a little general knowledge.
- **Hardware**: on Hopper GPUs (H100) flash-linear-attention refuses the Gated DeltaNet backward with torch 2.8's Triton
  (incorrect results, fla #640): train on Ampere/Ada (A100, L4). On Macs the DeltaNet layers run a slow PyTorch fallback.
- It answers the questions it is asked; it does not verify identity, and it must not be used to infer sensitive
  attributes (the training data asks apparent age only; gender and ethnicity were deliberately never asked).

## Licence

The adapter and head are derived from Kev-0.8B (Apache-2.0) and Qwen3.5 (Apache-2.0), but were trained on data that
includes research-only / non-commercial sources (RVL-CDIP, AVA, ScienceQA, AG News, Yelp; see the dataset card).
**Use for research only.** A commercially usable version would have to be retrained on permissively licensed sources only.