kai-decision / K-IM-Deiction

K-IM-Deiction V2 supplies seven heads for Chinese chat windows. Check its teacher-labelled results, local MLX port and training-data privacy limits.

By Songkai · First Hugging Face package release: 5 Nov 2026

A Qwen3.5-0.8B model with seven fixed heads for Chinese enterprise group chats. One pass labels category, sentiment, importance, urgency, intensity, attention and to-dos. K-IM-Deiction V2 ships merged weights, an optional adapter and an MLX port. The private chat data is not released.

Quick answers

Training data and privacy

V2 trained on 6,580 private Chinese group-chat windows. The maker says pseudonymous sender IDs accompanied verbatim message bodies sent to BigModel and DeepSeek APIs during labelling. Later body scrubbing of retained local data does not change the trained weights. Group names remain in local headers and some nicknames may remain; the card claims neither full anonymisation nor zero leakage. Local inference does not establish private training-data handling.

Source

Facts

Package

Merged Qwen3.5-0.8B V2 backbone, seven heads and calibration files; optional rank-64 adapter and native MLX edition

Inputs

Text windows, left-truncated to 512 tokens

Licence

Apache-2.0

Newer package

V2 weights uploaded 6 Nov 2026; MLX edition 7 Nov 2026

Claims and sources

• On 1,000 windows from seven unseen groups, the maker reports category macro-F1 0.611 across seven classes, or 0.673 after excluding classes with support below ten. Needs-attention AUC is 0.822; binary-head ECE is 0.053 and 0.069. Labels come from GLM-5.3, with a second teacher used for agreement checks. Median latency is 41.0 ms on an RTX 4090 D at batch one. The earlier DeepSeek comparison was not rerun for V2. vendor reportedsource

Links

Source: K-IM-Deiction card and merged weights → Source: Package history →

Related projects

No related projects are listed in the registry.

All projects · Back to the model registry