esco-gemma4

Live Interactive Demo: https://huggingface.co/spaces/mazafard/esco-gemma4-demo

This model is a fine-tuned version of gemma-4 designed to map candidate skills, job descriptions, and professional qualifications to official ESCO v1.2.1 (European Skills, Competences, Qualifications and Occupations) taxonomies and ISCO-08 unit codes.

Available Model Formats

📝 Prompt Formatting & Unsloth Usage

This model was trained using the standardized Alpaca prompt template. For optimal occupational classification, format your inputs as follows:

Prompt Template:

Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.

### Instruction:
Map the following professional skills and experience to the correct ESCO occupation title and ISCO-08 code.

### Input:
<ENTER_CANDIDATE_SKILLS_HERE>

### Response:

Complete Python Inference with Unsloth:

from unsloth import FastLanguageModel

# 1. Load fine-tuned 16-bit or 4-bit model
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="mazafard/esco-gemma4",
    max_seq_length=2048,
    dtype=None,
    load_in_4bit=True,
)
FastLanguageModel.for_inference(model)

# 2. Define the Alpaca prompt template
alpaca_prompt = """Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.

### Instruction:
Map the following professional skills and experience to the correct ESCO occupation title and ISCO-08 code.

### Input:
{}

### Response:
"""

# 3. Format input skills
skills = "Kubernetes, Docker container orchestration, Terraform IaC, CI/CD pipeline automation, Python, Prometheus monitoring, AWS cloud infrastructure"

inputs = tokenizer(
    [alpaca_prompt.format(skills)],
    return_tensors="pt"
).to("cuda")

# 4. Generate deterministic output
outputs = model.generate(
    **inputs,
    max_new_tokens=64,
    use_cache=True,
    temperature=0.1,
    do_sample=False
)

decoded = tokenizer.decode(outputs[0], skip_special_tokens=True)
response = decoded.split("### Response:")[-1].strip()
print("🎯 ESCO Classification Result:
", response)

Expected Model Output Format:

ESCO Occupation Title: Cloud Engineer / DevOps Specialist
ISCO-08 Code: 2512

⚡ Edge & Local Runtime Quickstart

Ollama (Desktop / CLI)

ollama run hf.co/mazafard/esco-gemma4-GGUF:q4_k_m

Apple Silicon Native MLX (macOS)

mlx_lm.generate \
    --model mazafard/esco-gemma4-MLX \
    --prompt "### Instruction:\nMap the following professional skills and experience to the correct ESCO occupation title and ISCO-08 code.\n\n### Input:\nKubernetes, Docker, CI/CD\n\n### Response:\n" \
    --max-tokens 64

Training Framework

• Base Model: Google Gemma 4 (4-bit QLoRA)
• Fine-Tuning Engine: Unsloth
• Dataset: European Commission ESCO Taxonomy v1.2.1

📄 Citation

To cite this project or the ESCO Taxonomy in academic research or applications, please use:

@misc{esco_taxonomy_2024,
  author       = {{European Commission}},
  title        = {European Skills, Competences, Qualifications and Occupations (ESCO) Dataset v1.2.1},
  year         = {2024},
  publisher    = {European Union Portal},
  howpublished = {\url{https://esco.ec.europa.eu/}},
  note         = {Accessed: 2026-05-26}
}

@software{gemma4_esco_finetune_2026,
  author       = {Fard, Mohammadreza A.},
  title        = {Parameter-Efficient Fine-Tuning (PEFT) and Telemetry Pipeline for Gemma-4 on ESCO Skill Inventories},
  month        = may,
  year         = {2026},
  publisher    = {GitHub Repository},
  version      = {1.0.0},
  url          = {https://github.com/mazafard/esco-gemma4-pipeline}
}