---
library_name: onnxruntime
tags:
  - onnx
  - int8
  - quantized
  - decision-making
  - wasm
  - browser-inference
  - encoder-only
base_model: jhu-clsp/ettin-encoder-68m
datasets:
  - avbiswas/bev-decision-150K
pipeline_tag: text-classification
license: apache-2.0
---

# Decider Ettin 68M

A 68M-parameter decision engine fine-tuned from [Ettin Encoder 68M](https://huggingface.co/jhu-clsp/ettin-encoder-68m) on the [BEV Decision 150K](https://huggingface.co/datasets/avbiswas/bev-decision-150K) dataset. Exported to ONNX (opset 18) and quantized to INT8 for browser inference via ONNX Runtime Web (WASM).

**Everything runs in the browser.** The model is downloaded once, cached in IndexedDB, and never requires a server for inference.

## Decision Heads

The model has three output heads activated by the input sequence structure:

| Head | Output | Description |
|------|--------|-------------|
| **Choice** | `choice_logits` | Pick one of up to 10 options |
| **Noul** | `noul_logits` | Binary yes / no verdict |
| **Score** | `score_logits` | Rate on a scale of up to 7 levels |

## Training Metrics

| Metric | Epoch 1 | Epoch 2 |
|--------|---------|---------|
| Choice accuracy | 62.5% | **65.8%** |
| Noul accuracy | 81.3% | **83.2%** |
| Score accuracy | 57.5% | **61.6%** |
| Score expected-value MAE | 0.727 | **0.645** |

Hyperparameters: encoder LR 1e-5, head LR 2.5e-4, max sequence length 2048.

## Input Sequence Format

The model expects a structured token sequence with trained special markers:

```
[DECISION] [STATE] <state text> [QUESTION] <question text> [OPT_0] <option 0> [OPT_1] <option 1> ...
```

Special tokens are pre-trained in the vocabulary at fixed IDs — they must **not** be added or recreated at runtime:

| Token | ID | Token | ID |
|-------|----|-------|----|
| `[DECISION]` | 50368 | `[OPT_0]` – `[OPT_9]` | 50371 – 50380 |
| `[STATE]` | 50369 | | |
| `[QUESTION]` | 50370 | | |

## ONNX Graph

**Inputs** (batch size 1, dynamic sequence length):

| Name | Type | Shape |
|------|------|-------|
| `input_ids` | int64 | `[1, sequence]` |
| `attention_mask` | int64 | `[1, sequence]` |
| `decision_pos` | int64 | `[1]` |
| `option_pos` | int64 | `[1, 10]` |
| `option_mask` | bool | `[1, 10]` |
| `score_n` | int64 | `[1]` |

**Outputs**:

| Name | Type | Description |
|------|------|-------------|
| `choice_logits` | float32 | Logits over option slots |
| `noul_logits` | float32 | `[no, yes]` logits |
| `score_logits` | float32 | Logits over score levels |

Unused heads output a masked sentinel (float32 minimum).

## Files

| File | Size | Description |
|------|------|-------------|
| `model_int8.onnx` | 67.3 MB | INT8 quantized ONNX model |
| `tokenizer.json` | 3.4 MB | BPE vocabulary (13 trained special tokens) |
| `tokenizer_config.json` | 613 B | Tokenizer settings |
| `config.json` | 1.1 KB | Training config, metrics, graph spec |
| `manifest.json` | 293 B | Version, size, SHA-256 for the web app |

**SHA-256** (`model_int8.onnx`): `cfd0184cf450c7ceeaab10ef1a7069bd8924ba46b0d9f73fd13a6a51d9616bd1`

## Usage

### Browser ([Decider web app](https://decider-ai.pages.dev/))

The companion web app downloads the model from this repo, caches it in IndexedDB, and runs inference entirely client-side:

```js
// Programmatic API exposed on window
await downloadModel();       // streams from HF, verifies SHA-256, saves to IndexedDB
await loadModel();           // builds ONNX Runtime WASM session

const result = await predict({
  type: "choice",
  state: "Alice has two projects and limited time.",
  question: "Which project should Alice choose?",
  options: ["Project A", "Project B"],
});

console.log(result.choice);  // logits → softmax for probabilities
```

### Python (reference)

```python
import onnxruntime as ort

session = ort.InferenceSession("model_int8.onnx")
print([inp.name for inp in session.get_inputs()])
# ['input_ids', 'attention_mask', 'decision_pos', 'option_pos', 'option_mask', 'score_n']
```

## Limitations

- Single-threaded WASM inference is slow on sequences above ~512 tokens.
- INT8 quantization trades some accuracy for a 4× size reduction over FP32.
- The model was trained on synthetic decision scenarios; real-world calibration has not been evaluated.

## License

Apache 2.0