---
license: apache-2.0
library_name: gguf
pipeline_tag: text-classification
base_model: mistralai/Ministral-3-3B-Instruct-2512
base-model: mistralai/Ministral-3-3B-Instruct-2512
language:
  - en
  - de
  - fr
  - es
  - multilingual
tags:
  - ollaya
  - ollama
  - trainaya
  - decision-model
  - system-1
  - llm-logits-v1
  - gguf
  - calibrated
  - ministral
datasets:
  - ZefanCai/Open-Jev-v1.1
  - SargeDev/jev-distill-corpus-v3
  - tasksource/tasksource-jev-typed-decisions
---

# push:3b — Ultra-Fast Calibrated System-1 Decision Model (3B GGUF)

**`push:3b`** (`solarpush/push-3b`) is an ultra-low-latency (`86 ms` p50, `2.31 GiB` on disk, `~4.2 GB` VRAM), calibrated **System-1 Decision Model** (`llm-logits-v1`) designed for single-forward-pass structured decision-making across three universal primitives:

- **`choice`** — Multi-way categorical selection (`A`..`Z`, 2 to 26 options),
- **`noul`** — Binary epistemic / policy / safety / entailment verification (`Yes` vs `No`),
- **`score`** — Ordinal rubric & severity evaluation (`0`..`9`, 3- to 5-level calibrated scales).

Built on a dense **3B parameter** GGUF backbone (`Q4_K_M`, `2.31 GiB`) with an untied **`F16` (`ttype=1`) subspace-orthogonal decision head** (`[131072, 3072]`) and cardinality-conditioned temperature scaling (`temperature_by_options`), `push:3b` achieves **`3.1%` Pooled Public ECE**, **`85.7%` accuracy on `civil_comments`**, **`84.0%` on `summeval-consistency`**, **`74.6%` on `squad2`**, and **`73.2%` on `pubmedqa`**.

---

## 1. Official Ollaya Decision Benchmark (`bench.ollaya.dev` — 14 Tasks, 5,880 Questions)

Evaluated strictly zero-shot / out-of-domain across all **14 official `bench.ollaya.dev` tasks** (`3,880` Public benchmark questions across 13 datasets + `2,000` `typed-decisions` workflow questions across 400 multi-question JSON states = **`5,880` total questions**, zero benchmark train/dev/test overlap):

| Model | Overall Skill `[95% CI]` | Public Skill *(13 Tasks)* | Public Macro Acc | Typed-Decisions Acc *(Skill)* | Public Pooled ECE *(lower is better)* | Latency *(p50)* | VRAM | License |
| :--- | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: |
| **`push:3b`** *(Adapted 3B `v12+`)* | **`46.0` `[43.9, 48.2]`** | **`47.1`** | **`64.8%`** | **`53.9%` (`32.5`)** | **`3.1%` (`0.031`)** | **`86 ms`** | **`4.2 GB`** | `Apache-2.0` |
| `push_ministral:3b` *(Baseline Adapted 3B)* | `44.1` `[42.0, 46.2]` | `45.3` | `63.7%` | `50.6%` (`27.6`) | `6.5%` (`0.065`) | `86 ms` | `4.2 GB` | `Apache-2.0` |
| **`push:26b`** *(26B-A4B MoE)* | **`62.7` `[60.8, 64.7]`** | **`63.1`** | **`75.4%`** | **`71.4%` (`58.1`)** | **`2.8%` (`0.028`)** | `223 ms` | `15.0 GB` | `Apache-2.0` |

---

## 2. Per-Task Breakdown (14 Tasks — 5,880 Questions)

| Benchmark Task | Primitive | Questions | `push:3b` (`v12+`) Accuracy | `push:3b` Skill *(0–100)* | `push_ministral:3b` Baseline *(Skill)* | Delta vs Baseline | `push:3b` ECE *(10-bin)* | Median Latency |
| :--- | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: |
| **`civil_comments`** *(Toxicity Detection)* | `noul` | 300 | **`85.7%`** | **`71.3`** | `83.7%` (`67.3`) | **`+2.0%` (`+4.0` Skill)** | **`0.043`** | `51 ms` |
| **`summeval-consistency`** *(Factual Consistency)* | `score` | 144 | **`84.0%`** | **`80.0`** | `83.3%` (`79.2`) | **`+0.7%` (`+0.8` Skill)** | `0.151` | `102 ms` |
| **`massive-en-US`** *(18-Way Intent)* | `choice` | 350 | **`76.0%`** | **`74.6`** | `75.7%` (`74.3`) | **`+0.3%` (`+0.3` Skill)** | **`0.064`** | `65 ms` |
| **`squad2`** *(Unanswerable QA)* | `noul` | 299 | **`74.6%`** | **`49.2`** | `75.6%` (`51.2`) | `-1.0%` | **`0.056`** | `67 ms` |
| **`boolq`** *(Reading Verification)* | `noul` | 300 | **`74.3%`** | **`48.7`** | `75.7%` (`51.3`) | `-1.4%` | **`0.039`** | `59 ms` |
| **`pubmedqa`** *(Biomedical 3-Way QA)* | `choice` | 250 | **`73.2%`** | **`59.8`** | `74.4%` (`61.6`) | `-1.2%` | **`0.069`** | `90 ms` |
| **`massive-de-DE`** *(18-Way Intent DE)* | `choice` | 350 | **`72.3%`** | **`70.7`** | `72.3%` (`70.7`) | `0.0%` | **`0.037`** | `67 ms` |
| **`aegis2`** *(AI Safety Guardrail)* | `noul` | 250 | **`64.8%`** | **`29.6`** | `68.4%` (`36.8`) | `-3.6%` | `0.125` | `54 ms` |
| **`multinli`** *(3-Way NLI)* | `choice` | 299 | **`62.2%`** | **`43.3`** | `55.5%` (`33.3`) | **`+6.7%` (`+10.0` Skill)** | `0.072` | `46 ms` |
| **`paws`** *(Adversarial Paraphrase)* | `noul` | 250 | **`60.0%`** | **`20.0`** | `61.2%` (`22.4`) | `-1.2%` | `0.105` | `50 ms` |
| **`vitaminc-dev`** *(3-Way Fact Verification)* | `choice` | 599 | **`58.1%`** | **`37.1`** | `52.9%` (`29.4`) | **`+5.2%` (`+7.7` Skill)** | **`0.048`** | `55 ms` |
| **`typed-decisions`** *(400 JSON Workflows)* | `mixed` | 2000 | **`53.9%`** | **`32.5`** | `50.6%` (`27.6`) | **`+3.3%` (`+4.9` Skill)** | `0.153` | `110 ms` |
| **`helpsteer2`** *(5-Level Helpfulness)* | `score` | 249 | **`33.7%`** | **`17.2`** | `35.3%` (`19.2`) | `-1.6%` | `0.127` | `109 ms` |
| **`summeval-relevance`** *(5-Level Relevance)* | `score` | 240 | **`24.2%`** | **`5.2`** | `13.8%` (`-7.8`) | **`+10.4%` (`+13.0` Skill)** | `0.175` | `124 ms` |
| **OVERALL (14 Tasks)** | **All** | **5880** | **`64.1%` (`64.8%` Public)** | **`46.0`** | `62.8%` (`44.1`) | **`+1.3%` (`+1.9` Skill)** | **`0.090` (`0.031` Pooled)** | **`86 ms`** |

### `typed-decisions` Enterprise Workflow Breakdown (2,000 Questions / 400 JSON States)

- **`security_incidents`**: **`64.0%`**
- **`agent_trace_observability`**: **`53.0%`** *(+6.4% vs baseline)*
- **`customer_service`**: **`52.8%`** *(+4.8% vs baseline)*
- **`invoice_processing`**: **`45.8%`** *(+2.4% vs baseline)*

---

## 3. Repository Contents

| File | Size | Description |
| :--- | :---: | :--- |
| `model.gguf` | `2.31 GiB` | 3B GGUF (`Q4_K_M` backbone + untied `F16` calibrated `output.weight` decision head) |
| `decision.json` | `2.5 KB` | Ollaya `llm-logits-v1` decision specification (`labels`, `template`, `temperature_by_options`, `llama` runtime config) |
| `calibration.json` | `751 B` | Post-hoc temperature scaling parameters and out-of-fold / post-quantization accuracy metrics |
| `ollaya_config.json` | `302 B` | Ollaya runtime manifest configuration |
| `Modelfile` | `559 B` | Deterministic single-token Ollama `Modelfile` (`temperature 0`, `top_k 1`, `num_predict 1`) |

---

## 4. Quickstart (Ollaya & Ollama)

### Option A: Register in both Ollaya and Ollama via `trainaya`

```bash
git clone https://huggingface.co/solarpush/push-3b
cd push-3b
trainaya register --model . --name push:3b --runtime both
```

### Option B: Upload to Hugging Face Hub from Trainaya Repo

```bash
make login
make push:3b
```