GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
seyhunak
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
seyhunak
Model Tree
Qwen/Qwen3.5-4B
Input Modalities
Output Modalities
Supported Functionality
Qwen/Qwen3.5-4B (configurable via --base-model).scripts/train.py).ANSWER: are masked (-100),
so cross-entropy focuses on the label. See README for rationale.state + question + options -> answer label.test.jsonl is never trained on.scripts/evaluate.py --mode base vs --mode tell on held-out test.jsonl.invalid_output as abstention.MIT (see LICENSE) for the Tell-4B adapter, code and synthetic starter data. Base-model weights follow their own license.
bash
pip install -r requirements.txtpython scripts/inference.py --adapter . --state "Invoice is 1200 EUR but PO is 1000 EUR." --question "What is the reconciliation result?" --options "A:match|B:mismatch|C:needs_review"# -> BA Q4_K_M GGUF of the merged model is published for local runs
(see Modelfile in the GitHub repo):
bash
ollama pull seyhunak/tell-4bollama run seyhunak/tell-4b "STATE: ... QUESTION: ... OPTIONS: ... ANSWER:"A Next.js one-pager classifying 50 finance tickets in real time
(demo/ in the GitHub repo, npx . after npm run build).
pip install -r requirements.txtpython scripts/inference.py --adapter . --state "Invoice is 1200 EUR but PO is 1000 EUR." --question "What is the reconciliation result?" --options "A:match|B:mismatch|C:needs_review"# -> Bollama pull seyhunak/tell-4bollama run seyhunak/tell-4b "STATE: ... QUESTION: ... OPTIONS: ... ANSWER:"