GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
hiteshluke
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
hiteshluke
Model Tree
unsloth/gemma-3-4b-it
Input Modalities
Output Modalities
Supported Functionality
| Benchmark | Primitive | Accuracy |
|---|---|---|
| JevBench boolq (n = 1,000) | noul (T / F) | 0.857 |
Independent 95 % CI on this run: [0.834, 0.877].
Classical "verbalizer" readout: the base model is asked a prompt that ends in
Answer:, and the probability mass over the single-token verbalizer set
(T, F) at the next position is read directly from the LM head. The LoRA
adapter sharpens this distribution without changing the architecture, so any
LLM pipeline that supports transformers + peft can serve this model
without a custom classification head.
This makes it:
transformers + peft run (CPU, CUDA, TGI, vLLM)python
from transformers import AutoTokenizer, AutoModelForCausalLMfrom peft import PeftModelimport torch BASE = "unsloth/gemma-3-4b-it"LORA = "hiteshluke/zyot-decider-v3-verbalizer-4b" tok = AutoTokenizer.from_pretrained(BASE)model = AutoModelForCausalLM.from_pretrained(BASE, dtype=torch.bfloat16, device_map="auto")model = PeftModel.from_pretrained(model, LORA).eval() T_ID = tok("T", add_special_tokens=False).input_ids[0]F_ID = tok("F", add_special_tokens=False).input_ids[0] def decide(passage: str, question: str) -> str: prompt = ( f"State: {passage}\n\n" f"Question: {question}\n\n" f"Options:\nT. Yes / True\nF. No / False\n\n" f"Answer:" ) ids = tok(prompt, return_tensors="pt", truncation=True, max_length=1024).input_ids.to(model.device) with torch.no_grad(): logits = model(ids).logits[0, -1] return "T" if logits[T_ID] > logits[F_ID] else "F" print(decide( "The Eiffel Tower is located in Paris, France.", "Is the Eiffel Tower in France?")) # -> Tadapter_config.json + adapter_model.safetensors — LoRA adapter (r = 16, α = 32) on Gemma 3 4Bunsloth/gemma-3-4b-it (text-only path)jev-distill-corpus-v3, noul-primitive rows at teacher-confidence ≥ 0.6T / F logitstransformers + peft — no custom server codeApache-2.0 for the LoRA adapter and code in this repository.
Base unsloth/gemma-3-4b-it and google/gemma-3-4b-it are governed by the Gemma license.
bibtex
@misc{zyot-decider-v3-verbalizer-2026, title = {Zyot Decider v3 — Verbalizer (4B)}, author = {Codekins Pvt Ltd · Zyot Lab}, year = {2026}, url = {https://huggingface.co/hiteshluke/zyot-decider-v3-verbalizer-4b}}from transformers import AutoTokenizer, AutoModelForCausalLMfrom peft import PeftModelimport torch BASE = "unsloth/gemma-3-4b-it"LORA = "hiteshluke/zyot-decider-v3-verbalizer-4b" tok = AutoTokenizer.from_pretrained(BASE)model = AutoModelForCausalLM.from_pretrained(BASE, dtype=torch.bfloat16, device_map="auto")model = PeftModel.from_pretrained(model, LORA).eval() T_ID = tok("T", add_special_tokens=False).input_ids[0]F_ID = tok("F", add_special_tokens=False).input_ids[0] def decide(passage: str, question: str) -> str: prompt = ( f"State: {passage}\n\n" f"Question: {question}\n\n" f"Options:\nT. Yes / True\nF. No / False\n\n" f"Answer:" ) ids = tok(prompt, return_tensors="pt", truncation=True, max_length=1024).input_ids.to(model.device) with torch.no_grad(): logits = model(ids).logits[0, -1] return "T" if logits[T_ID] > logits[F_ID] else "F" print(decide( "The Eiffel Tower is located in Paris, France.", "Is the Eiffel Tower in France?")) # -> T@misc{zyot-decider-v3-verbalizer-2026, title = {Zyot Decider v3 — Verbalizer (4B)}, author = {Codekins Pvt Ltd · Zyot Lab}, year = {2026}, url = {https://huggingface.co/hiteshluke/zyot-decider-v3-verbalizer-4b}}