QA2D-t5-small

This model is a fine-tuned version of t5-small on QA2D. It achieves the following results on the evaluation set:

• Loss: 0.3236
• Rouge1: 89.8753
• Rouge2: 81.8104
• Rougel: 85.4253
• Rougelsum: 85.4236
• Bleu: 72.1080

See: https://wandb.ai/domenicrosati/huggingface/runs/n1yallpe for training and eval stats and https://github.com/domenicrosati/qa2d-models for the code!

Model description

A t5-model model to convert questions, answer pairs into statements.

Due to the way it's been trained the input should be all lower case and punctuation removed. Use with . as the seperator between question and answer.

"where in the world is carmen. abruzzo" Output: "carmen is in abruzzo"

Thought punctation and upper case works.

from transformers import AutoTokenizer,  AutoModelForSeq2SeqLM

tokenizer = AutoTokenizer.from_pretrained('domenicrosati/QA2D-t5-small')
model = AutoModelForSeq2SeqLM.from_pretrained('domenicrosati/QA2D-t5-small')

question = "where in the world is carmen sandiego"
answer = "she is in abruzzo"
SEP = ". "

prompt = f'{question}{SEP}{answer}'
input_ids = tokenizer(prompt, return_tensors='pt').input_ids
output_ids = model.generate(input_ids)
responses = tokenizer.batch_decode(output_ids, skip_special_tokens=True)
# ['carmen sandiego is in abruzzo']

Intended uses & limitations

To convert questions, answer pairs into statements.

Training and evaluation data

Uses QA2D.

See https://github.com/domenicrosati/qa2d-models

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5.6e-05
• train_batch_size: 12
• eval_batch_size: 12
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 20
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Rouge1	Rouge2	Rougel	Rougelsum	Bleu
0.3177	1.0	5060	0.3144	89.6379	81.3168	85.2036	85.1904	71.4255
0.2479	2.0	10120	0.3035	89.7816	81.6556	85.3541	85.3406	71.7248
0.2268	3.0	15180	0.3015	89.8287	81.698	85.3434	85.3387	71.8344
0.2111	4.0	20240	0.3014	89.8082	81.7192	85.4094	85.406	71.9172
0.1991	5.0	25300	0.3023	89.8776	81.7607	85.3912	85.3842	71.9417
0.1886	6.0	30360	0.3012	89.901	81.7614	85.3345	85.3315	72.0218
0.1803	7.0	35420	0.3010	89.8776	81.8189	85.4154	85.4097	72.0533
0.1724	8.0	40480	0.3041	89.9168	81.8663	85.4457	85.4447	72.1470
0.1654	9.0	45540	0.3076	89.8901	81.8536	85.4857	85.4863	72.0830
0.1601	10.0	50600	0.3083	89.9186	81.881	85.4653	85.4594	72.1048
0.1546	11.0	55660	0.3136	89.8958	81.8533	85.4217	85.4238	72.0752
0.1502	12.0	60720	0.3138	89.903	81.8604	85.4301	85.4267	72.1373
0.1461	13.0	65780	0.3140	89.8867	81.7945	85.3698	85.3662	72.0718
0.1423	14.0	70840	0.3171	89.8985	81.8221	85.4348	85.4331	72.1168
0.1392	15.0	75900	0.3186	89.8938	81.8246	85.402	85.3991	72.0858
0.1366	16.0	80960	0.3208	89.859	81.8133	85.4194	85.4182	72.1014
0.1344	17.0	86020	0.3222	89.8909	81.828	85.4392	85.435	72.1380
0.1324	18.0	91080	0.3226	89.8906	81.8351	85.4506	85.4441	72.1622
0.1309	19.0	96140	0.3231	89.8925	81.8369	85.4375	85.4366	72.1552
0.1305	20.0	101200	0.3236	89.8753	81.8104	85.4253	85.4236	72.1080

Framework versions

• Transformers 4.18.0
• Pytorch 1.11.0+cu113
• Datasets 2.1.0
• Tokenizers 0.12.1