GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
Ephraimmm
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
Ephraimmm
Model Tree
unsloth/gemma-3-4b-it-unsloth-bnb-4bit
Input Modalities
Output Modalities
Supported Functionality
This is a Gemma-3-4B instruction-tuned model fine-tuned to handle customer service conversations for a Nigerian bank. The fine-tuning data reflects Nigerian customer-service interactions, so the model's responses are tuned to Nigerian tone and phrasing patterns as they appear in text (e.g. common Nigerian English expressions and code-switching seen in chat/support transcripts). It was trained with Unsloth and Hugging Face's TRL library using parameter-efficient (LoRA) fine-tuning.
| Base model | unsloth/gemma-3-4b-it-unsloth-bnb-4bit (Gemma 3, 4B, 4-bit quantized) |
| Method | Supervised fine-tuning with LoRA (via Unsloth + Hugging Face TRL) |
| LoRA rank (r) | 8 |
| LoRA alpha | 8 |
| LoRA dropout | 0 |
| Target modules | Attention projections (q_proj, k_proj, v_proj, o_proj) and MLP projections (gate_proj, up_proj, down_proj) of the language model |
| Task type | Causal LM |
| Precision | float16 |
This repository hosts the full fine-tuned model weights (merged, fp16, split across two safetensors shards) as well as the standalone LoRA adapter (adapter_model.safetensors) used to produce them, so the adapter can also be applied on top of the base model directly if preferred.
Training data, dataset size, and number of training steps/epochs are not recorded in this repository's metadata and are not published here.
This model is intended as a domain-specific assistant for Nigerian banking customer service scenarios rather than as a general-purpose chat model.
The repository contains the full merged model, so it can be loaded directly with transformers (no separate adapter step required):
python
import torchfrom transformers import AutoProcessor, Gemma3ForConditionalGeneration model_id = "Ephraimmm/customer-service" model = Gemma3ForConditionalGeneration.from_pretrained( model_id, device_map="auto", torch_dtype=torch.float16,)processor = AutoProcessor.from_pretrained(model_id) messages = [ {"role": "user", "content": [{"type": "text", "text": "Good day, please I want to check on my loan application status."}]}] inputs = processor.apply_chat_template( messages, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt",).to(model.device) with torch.no_grad(): output_ids = model.generate(**inputs, max_new_tokens=256) response = processor.decode( output_ids[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True)print(response)Alternatively, load with Unsloth for faster inference:
python
from unsloth import FastModel model, tokenizer = FastModel.from_pretrained( model_name="Ephraimmm/customer-service", max_seq_length=2048, load_in_4bit=True,)FastModel.for_inference(model)If you only want the LoRA weights, adapter_config.json / adapter_model.safetensors in this repo can be loaded on top of the base model unsloth/gemma-3-4b-it-unsloth-bnb-4bit with peft.
Developed by Ephraimmm
This gemma3 model was trained 2x faster with Unsloth and Hugging Face's TRL library.
import torchfrom transformers import AutoProcessor, Gemma3ForConditionalGeneration model_id = "Ephraimmm/customer-service" model = Gemma3ForConditionalGeneration.from_pretrained( model_id, device_map="auto", torch_dtype=torch.float16,)processor = AutoProcessor.from_pretrained(model_id) messages = [ {"role": "user", "content": [{"type": "text", "text": "Good day, please I want to check on my loan application status."}]}] inputs = processor.apply_chat_template( messages, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt",).to(model.device) with torch.no_grad(): output_ids = model.generate(**inputs, max_new_tokens=256) response = processor.decode( output_ids[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True)print(response)from unsloth import FastModel model, tokenizer = FastModel.from_pretrained( model_name="Ephraimmm/customer-service", max_seq_length=2048, load_in_4bit=True,)FastModel.for_inference(model)