wesley-stepgrid/hyperclovax-seed-text-instruct-1.5b-litert-lm

HyperCLOVA X SEED Text Instruct 1.5B — LiteRT-LM (safety fine-tuned, int4)

Powered by HyperCLOVA X.

This repository redistributes NAVER's HyperCLOVA X SEED Text Instruct 1.5B converted to the on-device .litertlm format (LiteRT-LM 0.13.x, Android), for use as the low-end Korean "mind" (노바 / Nova) in the StepGrid wellness app. It is a derivative of, and remains governed by, the HyperCLOVA X SEED Model License Agreement — see the license link above. All HyperCLOVA X SEED usage restrictions and the acceptable-use policy pass through to this redistribution unchanged.

What was changed vs. the base model

1. Safety fine-tuning (QLoRA). A small Korean-language safety SFT set (dangerous-question → safe-answer, with contrast/retention examples) was applied so the model reliably warns on household-chemical and gas-safety questions instead of giving unsafe advice. Merged back to the base weights before conversion.
2. Stop-token fix. eos_token_id set to the real turn-end tokens (<|im_end|>, <|endofturn|>) so generation stops cleanly (the base config.json value caused token leakage / run-on output).
3. int4 (BLOCKWISE_32 dynamic) quantization so it loads within memory on older Android CPUs.

The StepGrid app additionally applies deterministic, model-independent safety backstops (dangerous chemical-mix and medicine-dose guards) on top of this model at both generation and history-display time. On-device behavior should not be assumed safe without such guards.

Attribution

• Base model: HyperCLOVA X SEED Text Instruct 1.5B, © NAVER Cloud — Powered by HyperCLOVA X.
• License: HyperCLOVA X SEED Model License Agreement (see license_link). Redistribution here is under the same terms, including the name prefix, the "Powered by HyperCLOVA X" notice, and pass-through of the license and acceptable-use policy.