Gemma4-E2B-Function-Calling-xLAM-Unsloth

Gemma4-E2B-xLAM-Unsloth — Fine-tuned on xLAM-60k

Fine-tuned adapter-merged checkpoint of Gemma4-E2B-it on the Salesforce/xlam-function-calling-60k dataset, produced via Unsloth + Hugging Face TRL's SFTTrainer.

Usage

Python (transformers)

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "ermiaazarkhalili/Gemma4-E2B-Function-Calling-xLAM-Unsloth",
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
)
tokenizer = AutoTokenizer.from_pretrained("ermiaazarkhalili/Gemma4-E2B-Function-Calling-xLAM-Unsloth", trust_remote_code=True)

messages = [{"role": "user", "content": "Find flights from SFO to NYC on December 25th"}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text=text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=256, do_sample=False)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Unsloth (2× faster inference)

from unsloth import FastLanguageModel
model, tokenizer = FastLanguageModel.from_pretrained(
    "ermiaazarkhalili/Gemma4-E2B-Function-Calling-xLAM-Unsloth",
    max_seq_length=2048,
    load_in_4bit=False,
)
FastLanguageModel.for_inference(model)

GGUF (llama.cpp / Ollama)

Quantized GGUF versions are available at ermiaazarkhalili/Gemma4-E2B-Function-Calling-xLAM-Unsloth-GGUF:

# llama-cli
llama-cli -hf ermiaazarkhalili/Gemma4-E2B-Function-Calling-xLAM-Unsloth-GGUF --jinja -p "Find flights from SFO to NYC on December 25th" -n 256

# Ollama
ollama run hf.co/ermiaazarkhalili/Gemma4-E2B-Function-Calling-xLAM-Unsloth-GGUF:Q4_K_M

Training details

Function-calling fine-tuning on 60,000 examples from Salesforce/xlam-function-calling-60k.

• SFTTrainer (trl >= 0.14) via Unsloth's FastLanguageModel
• LoRA config: r=16, α=16, dropout=0, targeting q_proj/k_proj/v_proj/o_proj/gate_proj/up_proj/down_proj
• Effective batch size: 8 (per_device=2 × grad_accum=4)
• Max sequence length: 2048
• Optimizer: adamw_8bit with linear LR scheduler
• Seed: 3407

Intended use

For research and non-commercial experimentation only. Outputs should be independently verified before any downstream use.

Limitations

• Trained on a single epoch (~7,500 optimizer steps); further training may yield additional gains.
• Fine-tuned from Gemma4-E2B-it, inherits its limitations and biases.
• Evaluated only on training-data perplexity; no external benchmarks run on this checkpoint.
• Limited to the 60 function schemas covered in the training dataset; performance on novel APIs may degrade. Although the base model is a vision-language model (VLM), this fine-tune targets text-only function-calling; vision capabilities are untouched but unevaluated. GGUF sibling repo currently ships 3 of 6 planned quants (Q4_K_M, Q5_K_M, Q8_0); the 6-quant conversion exceeded memory limits on MIG 3g.40gb.

Citation

@misc{ gemma4_e2b_xlam_unsloth_2026 ,
  author = {Ermia Azarkhalili},
  title = { Gemma4-E2B-xLAM-Unsloth — Function-calling fine-tune of Gemma4-E2B-it },
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/ermiaazarkhalili/Gemma4-E2B-Function-Calling-xLAM-Unsloth}}
}

---

This gemma4 model was trained 2× faster with Unsloth and Hugging Face's TRL library.

<img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png" width="200"/>