pocket-voice-complete leg-1 — 303K params (dev checkpoint)

pocket-voice-complete leg-1: 141,729 → 302,977 params (2.14x: FFN 160→320, 2→3 transformer layers, third layer identity-initialized).

• 4 further epochs; eval accuracy 0.9845, eval loss 0.0402 — beats the base checkpoint's best (0.9800)
• Function preservation exact (max|diff| = 0.0)
• PyTorch transformer classifier for semantic completeness ('set a timer for five minutes' → 0.97 complete vs 'set a timer for' → 0.0006)
• Boundary: synthetic held-out accuracy only — the next meaningful leg needs real transcribed speech

Downloads:
https://github.com/ItsNotAILABS/pocket/raw/refs/tags/pocket-voice-complete-leg-1/checkpoints/pocket-voice-complete/leg-1/pytorch_model.bin
https://github.com/ItsNotAILABS/pocket/raw/refs/tags/pocket-voice-complete-leg-1/checkpoints/pocket-voice-complete/leg-1/receipt.json
https://github.com/ItsNotAILABS/pocket/raw/refs/tags/pocket-voice-complete-leg-1/checkpoints/pocket-voice-complete/leg-1/config.json

HIM-native leg-1 — 280K params (dev checkpoint)

HIM-native leg-1: 146,576 → 279,824 params (1.91x Net2Net widen, hidden 128→256).

• 800 steps on 2.26M chars of real text (Tiny Shakespeare, Alice, Sherlock, Frankenstein)
• Loss 5.5270 → 2.4063; preservation max|diff| = 8.3e-07
• Context-MLP causal LM (not a transformer); real NumPy manual-backprop/Adam trainer
• The v0 checkpoint was overfit on 4,109 tokens (train 0.08 / eval 3.54); this leg trains on real text to correct that

Real training on real text. Small model — not a capability claim.

Downloads:
https://github.com/ItsNotAILABS/pocket/raw/refs/tags/him-native-leg-1/checkpoints/him-native-v0/leg-1/weights.npz
https://github.com/ItsNotAILABS/pocket/raw/refs/tags/him-native-leg-1/checkpoints/him-native-v0/leg-1/receipt.json
https://github.com/ItsNotAILABS/pocket/raw/refs/tags/him-native-leg-1/checkpoints/him-native-v0/leg-1/config.json

auro.web.v1 leg-1 — 1.08M params (dev checkpoint)

auro.web.v1 leg-1: 557,696 → 1,081,984 params (1.94x Net2Net widen: attention heads 4→8, FFN 512→1024).

• 300 steps on 2.26M chars of real text (Tiny Shakespeare, Alice, Sherlock, Frankenstein)
• Loss 11.4035 → 3.1641; preservation max|diff| = 1.7e-05
• Byte-level causal transformer, NumPy handwritten autograd, real CE + AdamW trainer
• Note: head_dim (32) was made explicit in config post-training; tensor bytes unchanged, receipt re-hashed to match shipped bytes. Inference runtime needs explicit head_dim support for the widened (128,256) Q/K/V projections.

Real training on real text. Small byte-level model — not a capability claim.

Downloads:
https://github.com/ItsNotAILABS/pocket/raw/refs/tags/auro-web-v1-leg-1/vendor/auro_meaning/auro_web/leg-1/model.json
https://github.com/ItsNotAILABS/pocket/raw/refs/tags/auro-web-v1-leg-1/vendor/auro_meaning/auro_web/leg-1/receipt.json

POCKET 2.0.1-alpha

POCKET desktop + host source

• Windows (arm64 portable): attach below
• Host / operators: clone this repo and run Start-POCKET.ps1 or python -m pocket serve
• Live download page (if your tunnel is up): https://pocket.medinatechlabs.net/download

Paper/testnet-first for financial surfaces. Alpha quality.