mt5-small

This model is a fine-tuned version of google/mt5-small on an enhanced version of the Natural Questions dataset. It achieves the following results on the evaluation set:

• Loss: 0.7291
• Rouge1: 44.4366
• Rouge2: 38.8202
• Rougel: 43.113
• Rougelsum: 43.1423
• Bleu: 34.1596
• Gen Len: 12.6724
• Meteor: 0.4049
• True negatives: 69.7281
• False negatives: 10.4037
• Cosine Sim: 0.763

Model description

This model is fine-tuned for long-form, closed-domain question answering - question-answering from context. It uses a heavily refined version of Google's Natural Questions dataset.

Answers to the questions were rewritten using OpenAI's GPT-3.5 Turbo model.

Please see the following repo for all code and adaptations.

Intended uses & limitations

The model requires questions to be submitted using the following format using the input message: [CONTEXT] <\s> [QUESTION]

It is trained to respond appropriately when a question cannot be answered using the provided context.

It can give false negatives and false positives on occasion (see Training Results), and all answers must be checked appropriately.

Training and evaluation data

The model is trained using the Natural Questions dataset, with answers that have been refined using GPT-3.5 Turbo. It is evaluated using a number of metrics including BLEU, ROUGE, METEOR, and cosine similarity.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

# Load model and tokenizer
model_name = "psxjp5/mt5-small"
model = AutoModelForSeq2SeqLM.from_pretrained(model_name)
tokenizer = AutoTokenizer.from_pretrained(model_name)

# Generate text
context = "Once upon a time"
question = "What is time"
input_ids = tokenizer(context, question, return_tensors="pt").input_ids
outputs = model.generate(input_ids, max_new_tokens=150)

print(tokenizer.decode(output[0], skip_special_tokens=True))

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.001
• train_batch_size: 16
• eval_batch_size: 16
• seed: 9
• gradient_accumulation_steps: 8
• total_train_batch_size: 128
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 20
• weight_decay: 0.007
• dropout: 0.4

Training results

Training Loss	Epoch	Step	Validation Loss	Rouge1	Rouge2	Rougel	Rougelsum	Bleu	Gen Len	Meteor	True negatives	False negatives	Cosine Sim
2.5724	1.0	175	0.9876	18.7781	15.6002	18.22	18.2686	7.6676	7.7661	0.1628	72.8701	56.677	0.4003
1.1469	1.99	350	0.8580	36.8209	31.2514	35.5008	35.5462	25.7137	12.0014	0.3311	62.8399	20.3934	0.66
0.9468	2.99	525	0.7997	40.4128	34.716	39.0867	39.0972	29.3028	12.4287	0.3656	63.4441	15.295	0.7114
0.8129	3.98	700	0.7733	42.6764	36.7266	41.2465	41.2833	32.0644	12.9002	0.3871	62.1752	11.413	0.7425
0.7228	4.98	875	0.7483	42.9082	36.957	41.482	41.5233	32.4942	12.8866	0.3906	63.3233	11.5166	0.747
0.6493	5.97	1050	0.7293	40.3205	34.9632	39.1111	39.1168	28.8249	11.6867	0.3674	73.8973	17.9865	0.7068
0.5883	6.97	1225	0.7172	42.7342	37.0855	41.4069	41.424	32.1296	12.48	0.3887	70.0302	12.7847	0.7392
0.5409	7.96	1400	0.7387	44.6657	38.8426	43.3276	43.3496	34.4773	12.9395	0.4084	66.3444	9.5238	0.7658
0.5035	8.96	1575	0.7330	43.4925	38.0013	42.2697	42.2372	32.6131	12.2789	0.3979	72.6284	12.8364	0.7```1
0.4652	9.95	1750	0.7291	44.4366	38.8202	43.113	43.1423	34.1596	12.6724	0.4049	69.7281	10.4037	0.763

Framework versions

• Transformers 4.31.0
• Pytorch 2.0.1+cu118
• Datasets 2.13.1
• Tokenizers 0.13.3