qwen2.5-7b-parhaf-v1-synth-v2-xml-en-roles

LoRA adapter on unsloth/Qwen2.5-7B-Instruct for French clinical NER / de-identification (PARHAF schema), with role qualifiers on person/location/contact entities.

Trained on PARHAF_v1 (cleaned gold) + synthetic_v2 (FAMILY_STATUS re-annotated with marital status and kin/household mentions).

Intended use

• Extract PARHAF entity types from French clinical reports (XML tagging).
• Predict roles (Patient, Carer, Hospital, Other) when applicable.
• Research / benchmarking of clinical de-identification — not a medical device.

Training data

Split	Source	Docs
Train	PARHAF_v1 + synth_v2	504
Dev (eval)	PARHAF_v1	101

• PARHAF_v1: manually reviewed annotations (titles, ages-as-dates, FAMILY_STATUS, PSI, dates, CITY, etc.).
• synth_v2: same synthetic CRH corpus as v1, with FAMILY_STATUS re-labeled via Mistral to include kin (spouse, parents, children, …), aligned with the PARHAF_v1 definition.

Training setup

Hyperparameter	Value
Base model	unsloth/Qwen2.5-7B-Instruct
Method	LoRA (Unsloth / TRL SFT), r=16, alpha=16
Format	XML, English system prompt
Roles	yes (--with-roles)
Entity types	full (PARHAF ∪ AP-HP overlap, incl. FAMILY_STATUS)
Context	chunks 1500 chars, overlap 300
Epochs	3
Max seq length	4096

ft_config.json:

{
  "with_roles": true,
  "format": "xml",
  "prompt_lang": "en",
  "schema": "parhaf",
  "entity_types": "full",
  "context_mode": "chunk",
  "chunk_size": 1500,
  "overlap": 300
}

Evaluation (PARHAF_v1 /dev, offset, post-filtered)

Metric	F1
Entity (with label)	0.881
Entity (without label)	0.882
Char BIO (with label)	0.924
Role end-to-end	0.971
Role on matched entity spans	0.999

Selected per-label entity F1: FIRST_NAME 0.995, LAST_NAME 0.958, IDENTIFYING_DATE 0.940, FAMILY_STATUS 0.622, PATIENT_SOCIAL_IDENTITY 0.629.

Cross-eval on AP-HP (aphp2 test) not included in this card yet.

Entity labels

FIRST_NAME, LAST_NAME, IDENTIFYING_DATE, PATIENT_BIRTHDATE, PATIENT_SOCIAL_IDENTITY, FAMILY_STATUS, CITY, COUNTRY, ADDRESS, ZIP, MAIL, CODEIDENTIFIANT, PHONE_NUMBER, PATIENT_NATIONALITY, UNIDENTIFYING_DATE, URL (+ AP-HP overlap extras when present in mix).

FAMILY_STATUS covers marital status and identifying kin/household mentions (spouse, parents, children, siblings, etc.).

Load

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = "unsloth/Qwen2.5-7B-Instruct"
adapter = "niamaelkhbir/qwen2.5-7b-parhaf-v1-synth-v2-xml-en-roles"  # after push

tok = AutoTokenizer.from_pretrained(adapter)
model = AutoModelForCausalLM.from_pretrained(base, device_map="auto", torch_dtype="auto")
model = PeftModel.from_pretrained(model, adapter)

Related adapters

• niamaelkhbir/qwen2.5-7b-parhaf-v1-xml-en — PARHAF_v1 only (no synth, no roles)
• niamaelkhbir/qwen2.5-7b-parhaf-xml-en-chunk — earlier PARHAF (pre-cleanup)

Framework versions

• PEFT 0.19.1
• TRL 0.24.0
• Transformers 5.5.0
• PyTorch 2.10.0
• Unsloth

Citation

If you use this adapter, please cite the PARHAF / CU1 clinical NER work and:

@misc{vonwerra2022trl,
  title        = {{TRL: Transformer Reinforcement Learning}},
  author       = {Leandro von Werra and others},
  year         = 2020,
  howpublished = {\\url{https://github.com/huggingface/trl}}
}