GovLLM-7B-ultra

A question answering model about the Dutch Government.

Model description

This model is a fine-tuned version of the Dutch conversational model BramVanroy/GEITje-7B-ULTRA on a Dutch question-answer pair dataset of the Dutch Government. This is a Dutch question/answer model ultimately based on Mistral and fine-tuned with SFT and LoRA. The training with 3 epochs took almost 2 hours and was run on an Nvidia A100 (40GB VRAM).

Usage with Inference Endpoints (Dedicated)

import requests

API_URL = "https://your-own-endpoint.us-east-1.aws.endpoints.huggingface.cloud"
headers = {"Authorization": "Bearer hf_your_own_token"}

def query(payload):
    response = requests.post(API_URL, headers=headers, json=payload)
    return response.json()

output = query({
    "inputs": "Geeft de overheid subsidie aan bedrijven?"
})

# print generated answer
print(output[0]['generated_text'])

Training hyperparameters

The following hyperparameters were used during training:

• block_size: 1024,
• model_max_length: 2048,
• padding: right,
• mixed_precision: fp16,
• learning rate (lr): 0.00003,
• epochs: 3,
• batch_size: 2,
• optimizer: adamw_torch,
• schedular: linear,
• quantization: int8,
• peft: true,
• lora_r: 16,
• lora_alpha: 16,
• lora_dropout: 0.05

Training results

Epoch	Loss	Grad_norm	learning_rate	step
0.14	1.3183	0.6038	1.3888e-05	25/540
0.42	1.0220	0.4180	2.8765e-05	75/540
0.69	0.9251	0.4119	2.56793-05	125/540
0.97	0.9260	0.4682	2.2592e-05	175/540
1.25	0.8586	0.5338	1.9506e-05	225/540
1.53	0.8767	0.6359	1.6420e-05	275/540
1.80	0.8721	0.6137	1.3333e-05	325/540
2.08	0.8469	0.7310	1.0247e-05	375/540
2.36	0.8324	0.7945	7.1605e-05	425/540
2.64	0.8170	0.8522	4.0741e-05	475/540
2.91	0.8185	0.8562	9.8765e-05	525/540