GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
Zeyad
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
Zeyad
Model Tree
humain-ai/ALLaM-7B-Instruct-preview
Input Modalities
Output Modalities
Supported Functionality
QLoRA NF4; rank 16; alpha 32; dropout 0.05; learning rate 5e-05; maximum 1.0 epoch; maximum sequence 2048; response-token mean loss per example; microbatch 1 and gradient accumulation 16; no packing. Dataset sizes: {'train': 5630, 'validation': 736, 'test': 695}. Data are curated social-media post/comment pairs. Comments from an original post and detected lexical duplicate prompts stay in one split. Data selection uses manual exclusion marks and a cap of 5 comments per post, ordered by existing likes rank.
Validation loss selects a checkpoint. This alone does not prove improved social quality, cultural representativeness or preservation of general capabilities. Blind baseline-versus-adapter response review is provided in the training notebook; no human win-rate or benchmark improvement is claimed here. The corpus is concentrated in one Reddit community. Saudi author identity is not verified. Not established as representative of all Saudi dialects or populations. May hallucinate or imitate personal claims. General multilingual/reasoning retention has not been evaluated. Not a professional advice system. Raw training and evaluation rows are not included in this repository.
See training_recipe.json and environment.txt. The adapter requires the base weights; it is not a standalone merged model.