perspectief-llama-3-8b-dutch-qa-lora-v1

base_model: unsloth/meta-llama-3.1-8b-instruct-bnb-4bit

Model Card for perspectief-llama3-8b-adapter

A parameter-efficient LoRA adapter that turns Meta-Llama-3.1-8B-Instruct into a Dutch‑centric, context‑aware assistant for questions about Perspectief (occupational health, safety and absenteeism services).

The adapter was trained on ≈ 400 Q/A rows automatically extracted from perspectief and formatted as:

<|user|>
{Question}

Context:
{DetailedContext}
<|assistant|>
{Answer}<|end|>

Its main purpose is to be used in a Retrieval‑Augmented Generation (RAG) pipeline together with a FAISS knowledge base built from the same source material.

---

Model Details

Model Sources

• Repository: to be published – https://github.com/ org/perspectief-llama3-rag

---

Intended Uses

• Direct use – Drop‑in adapter for Llama‑3.1‑8B to answer Perspectief‑specific questions when a context passage is supplied.
• Out‑of‑scope – Open‑domain Q/A, creative writing, or usage without an external context (hallucination risk).

---

Bias, Risks & Limitations

Content reflects websites as of May 2025; anything outside that scope may be outdated or wrong. Always display the retrieved passage so users can verify.

---

Quick Start

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

BASE = "meta-llama/Meta-Llama-3.1-8B-Instruct"
ADAPT = "checkpoints/perspectief-llama3"  # adapter folder

tok = AutoTokenizer.from_pretrained(BASE, use_fast=True)
tok.add_special_tokens({"additional_special_tokens": ["<|user|>","<|assistant|>","<|end|>"]})

base = AutoModelForCausalLM.from_pretrained(BASE, device_map="auto")
model = PeftModel.from_pretrained(base, ADAPT)

prompt = """<|user|>\nWat kost een extra collega bij de 4‑uur durende training?\n\nContext:\nThe €595 fee applies to each primary participant…\n<|assistant|>\n"""

print(tok.decode(model.generate(**tok(prompt, return_tensors="pt").to(model.device),
                                max_new_tokens=128)[0]))

---

Training Summary

Training data: perspectief_dataset‑1.jsonl (417 rows).

---

Evaluation

Manual review on 20 held‑out questions shows ≥ 95 % exact‑match when combined with the FAISS retriever.

---

Contact

Open an issue in the GitHub repo for questions or feedback.

---

Framework Versions

• PEFT 0.15.2
• Transformers 4.41
• Unsloth 2025.5