Vini Pico — 49.9M Edge Personal Assistant
A 49.9M-parameter personal assistant trained from scratch using BitNet b1.58 ternary weights. Designed for everyday tasks, IoT control, and tool use on low-cost edge hardware.
What It Does
- Everyday tasks: reminders, timers, Q&A, summaries
- Smart home / IoT: sensors, lights, thermostat, locks, device control
- Tool routing: calling the right tool instead of hallucinating
- Friendly conversation: natural, helpful responses
- Safety: refuses unsafe requests, admits when it doesn't know
- On-device: runs locally on low-cost edge hardware, no cloud needed
Model Overview
| Property | Value |
|---|
| Parameters | 49.9M (ternary: -1, 0, +1) |
| Architecture | BitNet b1.58 |
| Layers | 16 |
| Dimension | 512 |
| Attention heads | 8 (2 KV, GQA) |
| FFN | ReLU² with SubLN |
| Context length | 2,048 |
| Vocabulary | 32K BPE |
| Quantization | 1.58-bit (packed ternary) |
| Target hardware | Raspberry Pi 4, old phones, edge devices |
Training Status
| Phase | Status | Details |
|---|
| Pretrain | Done | 7,972 steps, 1.045B tokens, best loss 4.48 at step 7500 |
| SFT iter 1–4 | Done | 17 bugs fixed, code/data issues resolved |
| SFT iter 5 | Pending | Fixed SFT dataset is ready; rerun is blocked on the planned credit refresh |
| DPO | Conditional | Only if SFT quality gates pass |
Dataset State
The current canonical dataset for Pico is the content-verified merged snapshot under:
vini/models/pico/data/generated/runs/run_20260524_1322/merged/dataset.jsonl
Current canonical snapshot:
- 5,965 records
- chat-style JSONL
- one top-level key per record:
messages
- verified metadata stored beside the dataset file
This corpus is a small-assistant curriculum, not a tool-only corpus and not a general coding corpus.
Repository Contents
Model artifacts: configurations, checkpoints, tokenizer, and exports.
| Path | Contents |
|---|
pico_config.yaml | Pico architecture config |
shared/ | Training configs (pretrain, SFT, DPO), tokenizer |
checkpoints/ | Future: model weights (.pt, .safetensors) |
exports/ | Future: .gguf, .onnx, .mlpackage |
Companion Repositories
Architecture
- BitNet b1.58: Ternary weights (-1, 0, +1) — efficient inference for edge deployment
- ReLU² FFN: strong sparsity and good fit for ternary weights
- SubLN: Prevents dead layers during training
- GQA: Grouped-Query Attention (8 heads, 2 KV)
- RoPE: Rotary Position Embeddings
Training Configuration
- Optimizer: AdamW (β1=0.9, β2=0.95), BF16 mixed precision
- LR Schedule: Two-stage cosine (4e-3 → 4e-4 cooldown → min_lr fix in training code)
- Batch: 64 effective (BS=4, GA=16)
- Framework: PyTorch with Straight-Through Estimator (STE)
Model Ladder
| Model | Params | Target Hardware | Status |
|---|
| Pico | 49.9M | Raspberry Pi 4 | Active |
| Nano | ~100M | Raspberry Pi 5 | Planned |
| Micro | ~500M | Modern phones | Planned |
| Mini | ~1B | Desktop | Planned |
License
Apache 2.0
