SKILL-Bridge

SKILL-Bridge

SKILL-Bridge is a domain-specialised language model that maps unstructured labour-market queries — job titles, hiring profiles, degree tracks, and academic modules — into validated JSON profiles annotated with ESCO v1.2.1 human-readable skill labels.

The model is designed for deterministic downstream integration: workflow engines (n8n), RAG platforms (AnythingLLM), and labour-market analytics pipelines that require machine-parseable skill inventories without ESCO URIs, UUIDs, or free-form reasoning traces.

Repository: hicham-taoufik/SKILL-Bridge
Architecture: Full merged weights — QLoRA adapter fused into Qwen/Qwen3.5-4B. No PEFT required at inference.

---

Abstract

Automated skill profiling from natural language remains challenging when outputs must conform to a rigid schema, span multiple languages, and reference a controlled vocabulary (ESCO) without leaking identifiers. General-purpose instruction models frequently emit markdown wrappers, chain-of-thought prose, or malformed JSON — particularly on long French degree descriptions and Arabic job postings.

SKILL-Bridge addresses this through supervised fine-tuning on 10,132 curated ChatML examples covering English, French, and Arabic inputs across four task types. Built on the instruction-aligned Qwen3.5-4B backbone with QLoRA (rank 64) and assistant-only loss on schema-constrained JSON, the model achieves 89.1% JSON validity on a 55-prompt held-out benchmark spanning occupations, job titles, degrees, and modules — enabling reliable structured skill extraction for production pipelines.

---

Model description

Task formulation

Given a user query describing a labour-market entity, the model produces a structured skill profile:

$$\text{query} \rightarrow {\texttt{resolved_title}, \texttt{essential_skills}, \texttt{optional_skills}, \ldots}$$

Input types (input_type):

Languages (language): en, fr, ar, or mixed.

---

Training data

Language distribution (train): EN 2,435 · FR 2,476 · AR 2,417

Input-type distribution (train): job_title 2,941 · target_occupation 2,554 · degree 1,112 · module 721

Each example is a ChatML conversation:

• System: ESCO skill advisor instructions (labels only, no URIs)
• User: Structured headers (input_type, language, text)
• Assistant: Validated JSON conforming to the UCB schema

The dataset includes ~534 negative examples (needs_review: true) and 1,000 intentionally messy free-text rows to improve robustness on real-world inputs.

---

Training procedure

Training run:

• Steps: 1,016 · Runtime: ~5 h 39 min
• Final train loss: 0.70 · Token accuracy: 90.7%
• Final step loss: 0.395

The LoRA adapter was merged into base weights with PeftModel.merge_and_unload() to produce this standalone checkpoint (~7.9 GB, 2 safetensor shards).

---

Evaluation

Benchmark: 55 diverse prompts (EN/FR/AR × occupation / job / degree / module), evaluated with temperature 0, JSON-prefix continuation, and schema validation.

SKILL-Bridge produces parseable, schema-conformant JSON across multilingual labour-market queries while maintaining ESCO label fidelity. Residual failure modes include occasional truncation on very long degree descriptions and rare category-field formatting edge cases.

Median inference latency (local, single GPU): 10–14 s per request for degree/job prompts at 896–1152 max new tokens.

---

Output schema

{
  "language": "fr",
  "input_type": "job_title",
  "resolved_title": "chef de projet informatique",
  "essential_skills": ["..."],
  "optional_skills": ["..."],
  "skill_categories": {
    "digital": ["..."],
    "transversal": ["..."],
    "green": ["..."]
  },
  "related_occupations": ["..."],
  "metrics": {
    "essential_count": 24,
    "optional_count": 18
  },
  "needs_review": false
}

Constraints enforced at training time:

• Maximum 30 essential + 30 optional skills per profile
• Human-readable ESCO labels only — no conceptUri, esco_id, or HTTP links
• Single JSON object — no markdown fences or explanatory prose

---

Usage

Transformers (recommended)

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

MODEL = "hicham-taoufik/SKILL-Bridge"

tokenizer = AutoTokenizer.from_pretrained(MODEL, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    MODEL,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
)
model.eval()

user_content = """input_type: job_title
language: fr
text: Nous recrutons un chef de projet informatique. Listez les compétences essentielles et optionnelles."""

messages = [
    {
        "role": "system",
        "content": (
            "You are an ESCO v1.2.1 labor-market skill advisor. "
            "Given a job title, hiring profile, degree track, or academic module, "
            "return required skills as human-readable labels only. "
            "Never output URIs, UUIDs, esco_id, conceptUri, or http links. "
            "Output one valid JSON object only — no markdown, no thinking, no commentary."
        ),
    },
    {"role": "user", "content": user_content},
    {"role": "assistant", "content": '{"language":"fr","input_type":"job_title"'},
]

prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=False,
    continue_final_message=True,
    chat_template_kwargs={"enable_thinking": False},
)

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
with torch.inference_mode():
    output_ids = model.generate(
        **inputs,
        max_new_tokens=1536,
        do_sample=False,
        temperature=0.0,
        repetition_penalty=1.12,
    )

response = tokenizer.decode(
    output_ids[0, inputs["input_ids"].shape[1]:],
    skip_special_tokens=True,
)
print(response)

OpenAI-compatible API

Deploy with any OpenAI-compatible server (vLLM, llama.cpp, custom FastAPI). Recommended generation settings:

Use structured user headers (input_type / language / text) for best results.

---

Limitations

• Skill labels are model-generated approximations of ESCO terminology; always validate against the official ESCO portal for compliance-critical applications.
• Long degree descriptions may occasionally truncate before JSON closure — increase max_new_tokens or post-process with a JSON repair utility.
• The model does not perform skill-to-occupation linking via URIs; downstream systems must resolve labels to ESCO concepts independently if identifiers are required.
• Performance on languages outside EN/FR/AR is not evaluated.
• Not suitable for open-ended chat; outputs are optimised for structured skill profiling only.

---

Ethical considerations

This model is intended for labour-market analytics, career guidance, and HR decision support — not for automated hiring decisions without human review. Outputs should be treated as advisory. The needs_review flag in the schema indicates low-confidence profiles that warrant manual inspection.

---

License

• Model weights: Apache 2.0
• Base model (Qwen3.5-4B): Subject to Qwen license terms
• ESCO skill labels: Subject to ESCO reuse policy

---

Citation

If you use SKILL-Bridge in academic or industrial research, please cite:

@misc{skillbridge2026,
  title        = {SKILL-Bridge: Multilingual ESCO Skill Profiling with Structured JSON Output},
  author       = {Taoufik, Hicham},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/hicham-taoufik/SKILL-Bridge}},
  note         = {Fine-tuned Qwen3.5-4B for ESCO v1.2.1 label-only skill extraction (EN/FR/AR)}
}

---

Acknowledgements

Built on Qwen3.5-4B (Alibaba Cloud). Skill taxonomy from the European Commission's ESCO classification. Training infrastructure: NVIDIA DGX Spark (GB10).