Modelpublic

Thox-ai/ThoxMini-125M-role

sourceHugging Faceapache-2.0updated 2mo agoView on Hugging Face

0likes226downloads

Model Card

ThoxMini-125M-role

The on-device model for ThoxMini (Pi Zero 2 W, arm64/NEON, 512 MB).

Thox-ai/thox-micro-125m role-tuned so it can actually be used as an assistant.

Why this exists

The base is a base LM with no instruct tuning. Measured on the exact file that had been dropped to the image lane:

Hello, who are you? → <mup> PR snapd#7621 closed: tests: check that the test fails with an error <Created by mvo5>

That is raw training-scrape leaking. It could not ship as an assistant. After tuning, the same prompt returns a coherent deferral.

Measured behaviour

Prompts held out of the training corpus:

prompt	answer
who funded thox	I do not have that on the device. Ask the host.
where is the headquarters	I do not have that on the device. Ask the host.
how much did the company raise	I do not have that on the device. Ask the host.
what is the weather	I do not have that on the device. Ask the host.

Zero fabrications, and no scrape leakage. In-corpus commands work: status → "Key online. Storage sealed. No pending transfers."

Two known defects — read before shipping

1. It says "the key", not "the mini." The role corpus was authored for ThoxKey-S3 and reused verbatim, so device-specific strings are wrong for this board: do you send my data anywhere → "No. I run on the key. Nothing leaves it." Correct in substance, wrong in noun. Fix is a ThoxMini variant of datasets/device/thoxkey-role.jsonl, not a retrain of the recipe.

2. It over-defers on identity. Hello, who are you? returns "I cannot answer that locally. The host has it." It should name itself. The corpus trains deferral hard and contains no identity turn for this device; adding one fixes it.

Neither is a fabrication, which is the failure mode that actually blocks shipping. Both are corpus gaps.

Training


base	`Thox-ai/thox-micro-125m` (Llama arch, 12 layers, vocab 50304)
corpus	1,360 device-role turns + 205 v5 chat turns = 1,565 rows
recipe	full fine-tune, 3 epochs, lr 5e-5, bf16, max_len 512
hardware	local RTX 4060 Ti, ~1 min, $0
final	loss ~0.90, mean token accuracy 0.86

v5 and not v4 deliberately. v4 carries tag: correction rows that quote a falsehood so the assistant can deny it; those rows taught fabrication rather than preventing it — a 7B trained on them invented "a Partner at Insight Partners" for a real named person. v5 quarantines them.

EOS is appended explicitly during training. Omitting it is a defect this factory already paid for: adapters trained without it emitted EOS zero times across 12 gate answers and ran to the token cap every time.