Instructions to use rubenbalbastre/procurement-function-calling-gemma-4-E2B-it-sft with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use rubenbalbastre/procurement-function-calling-gemma-4-E2B-it-sft with PEFT:

from peft import PeftModel
from transformers import AutoModelForCausalLM

base_model = AutoModelForCausalLM.from_pretrained("google/gemma-4-E2B-it")
model = PeftModel.from_pretrained(base_model, "rubenbalbastre/procurement-function-calling-gemma-4-E2B-it-sft")

Transformers

How to use rubenbalbastre/procurement-function-calling-gemma-4-E2B-it-sft with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="rubenbalbastre/procurement-function-calling-gemma-4-E2B-it-sft")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("rubenbalbastre/procurement-function-calling-gemma-4-E2B-it-sft", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use rubenbalbastre/procurement-function-calling-gemma-4-E2B-it-sft with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "rubenbalbastre/procurement-function-calling-gemma-4-E2B-it-sft"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "rubenbalbastre/procurement-function-calling-gemma-4-E2B-it-sft",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/rubenbalbastre/procurement-function-calling-gemma-4-E2B-it-sft

SGLang

How to use rubenbalbastre/procurement-function-calling-gemma-4-E2B-it-sft with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "rubenbalbastre/procurement-function-calling-gemma-4-E2B-it-sft" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "rubenbalbastre/procurement-function-calling-gemma-4-E2B-it-sft",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "rubenbalbastre/procurement-function-calling-gemma-4-E2B-it-sft" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "rubenbalbastre/procurement-function-calling-gemma-4-E2B-it-sft",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use rubenbalbastre/procurement-function-calling-gemma-4-E2B-it-sft with Docker Model Runner:
```
docker model run hf.co/rubenbalbastre/procurement-function-calling-gemma-4-E2B-it-sft
```

Procurement Function Calling — Gemma 4 E2B LoRA SFT

This repository contains a LoRA adapter for google/gemma-4-E2B-it, supervised fine-tuned for multilingual, multi-turn procurement tool calling. The model searches suppliers, requests quotes, evaluates delivery options and submits a procurement plan while respecting budget, delivery, reliability and compliance constraints.

Results

Every run evaluates all 400 scenario–prompt pairs in the held-out test split, including unseen templates 16–25. All configurations use the same seed and test examples. Values are the mean ± sample standard deviation across three repeated inference runs.

Model	Thinking	Runs	Task success	Average return	Mean episode latency
Base	Disabled	3	16.58% ± 0.29 pp	0.230 ± 0.010	15.75 ± 0.14 s
Base	Enabled	3	40.33% ± 1.66 pp	0.476 ± 0.017	95.36 ± 2.04 s
SFT	Disabled	3	42.50% ± 0.35 pp	0.567 ± 0.005	20.31 ± 0.19 s

Against the non-thinking base, SFT improves mean task success by 25.92 percentage points and mean return by 0.338. It also slightly exceeds the thinking-enabled base while using about one fifth of its mean episode latency.

The 48-step job processed 1,472,140 non-padding input tokens across 384 examples. Of these, 188,297 assistant tool-call tokens carried loss. Training took 65 minutes on one NVIDIA A40 GPU. The run was deliberately compute-bounded and was not optimized for maximum environment return or task success.

Loading

from peft import AutoPeftModelForCausalLM
from transformers import AutoTokenizer

model_id = "rubenbalbastre/procurement-function-calling-gemma-4-E2B-it-sft"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoPeftModelForCausalLM.from_pretrained(
    model_id,
    device_map="auto",
    torch_dtype="auto",
)

This repository contains the adapter rather than the complete base-model weights. Use the tokenizer's chat template and provide the procurement tool schemas when constructing requests.

Training

Base model: google/gemma-4-E2B-it
Dataset: rubenbalbastre/supply-chain-tool-calling
Method: assistant-only LoRA supervised fine-tuning with PEFT and TRL
LoRA rank: 16
LoRA alpha: 32
LoRA dropout: 0.05
Optimizer steps: 48
Thinking during SFT: disabled

Limitations

Results are specific to the included synthetic procurement environment, tools and verifier.
Preferred-supplier fallback remains the weakest task family.
The model can emit invalid calls, select infeasible plans or fail to complete longer trajectories.
Do not use its output as an autonomous purchasing decision without external validation.

Source

Training, data-generation and evaluation code is available in the sft-tool-calling repository.

Downloads last month: 29

Model tree for rubenbalbastre/procurement-function-calling-gemma-4-E2B-it-sft

Base model

google/gemma-4-E2B

Finetuned

google/gemma-4-E2B-it

Adapter

(199)

this model

Dataset used to train rubenbalbastre/procurement-function-calling-gemma-4-E2B-it-sft

Collection including rubenbalbastre/procurement-function-calling-gemma-4-E2B-it-sft

2B Tool Calling using SFT

Collection

https://github.com/rubenbalbastre/sft-tool-calling • 2 items • Updated 3 days ago