Model Card for Backpack-GPT2-NLI

This is a fine-tuned version of backpack-gpt2 with a NLI classification head on the esnli dataset. Results:

• On Validation Set:
  • CrossEntropyLoss: 0.3168
  • Accuracy: 0.9006
  • F1: 0.9004
• On Test Set:
  • CrossEntropyLoss: 0.3277
  • Accuracy: 0.8958
  • F1: 0.8955

Model Description

• Developed by: Erfan Moosavi Monazzah
• Model type: Sequence Classifier
• Language(s) (NLP): English
• License: apache-2.0
• Finetuned from model [optional]: Backpack-GPT2

How to Get Started with the Model

from transformers import AutoTokenizer, AutoModelForSequenceClassification

tokenizer = AutoTokenizer.from_pretrained('gpt2')
tokenizer.pad_token = tokenizer.eos_token

def tokenize_function(examples):
    concatenated_sentences = [f'{premise.strip(".")}. ^ {hypothesis.strip(".")}.' for premise, hypothesis in zip(examples['premise'], examples['hypothesis'])]

    tokenized_inputs = tokenizer(
        concatenated_sentences,
        padding="max_length",
        truncation=True,
        max_length=512,
        return_tensors="pt",
    )
    return tokenized_inputs

model = AutoModelForSequenceClassification.from_pretrained('ErfanMoosaviMonazzah/backpack-gpt2-nli', trust_remote_code=True)
model.eval()

tokenized_sent = tokenize_function({
    'premise':['A boy is jumping on skateboard in the middle of a red bridge.',
               'Two women who just had lunch hugging and saying goodbye.',
               'Children smiling and waving at camera'],
    'hypothesis':['The boy does a skateboarding trick.',
                  'The friends have just met for the first time in 20 years, and have had a great time catching up.',
                  'The kids are frowning']
})
model.predict(input_ids=tokenized_sent['input_ids'], attention_mask=tokenized_sent['attention_mask'])

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e-5
• train_batch_size: 64
• eval_batch_size: 64
• seed: 2023
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_ratio: 0
• num_epochs: 3

Training results

Step	Training Loss	Validation Loss	Precision	Recall	F1	Accuracy
512	0.614900	0.463713	0.826792	0.824639	0.825133	0.824731
1024	0.503300	0.431796	0.844831	0.839414	0.839980	0.839565
1536	0.475600	0.400771	0.848741	0.847009	0.846287	0.847795
2048	0.455900	0.375981	0.859064	0.857357	0.857749	0.857448
2560	0.440400	0.365537	0.862000	0.862078	0.861917	0.862426
3072	0.433100	0.365180	0.864717	0.859693	0.860237	0.859785
3584	0.425100	0.346340	0.872312	0.870635	0.870865	0.870961
4096	0.413300	0.343761	0.873606	0.873046	0.873174	0.873298
4608	0.412000	0.344890	0.882609	0.882120	0.882255	0.882341
5120	0.402600	0.336744	0.876463	0.875629	0.875827	0.875737
5632	0.390600	0.323248	0.882598	0.880779	0.881129	0.880817
6144	0.388300	0.338029	0.877255	0.877041	0.877126	0.877261
6656	0.390800	0.333301	0.876357	0.876362	0.875965	0.876753
7168	0.383800	0.328297	0.883593	0.883675	0.883629	0.883967
7680	0.380800	0.331854	0.882362	0.880373	0.880764	0.880512
8192	0.368400	0.323076	0.881730	0.881378	0.881419	0.881528
8704	0.367000	0.313959	0.889204	0.889047	0.889053	0.889352
9216	0.315600	0.333637	0.885518	0.883965	0.884266	0.883967
9728	0.303100	0.319416	0.888667	0.888092	0.888256	0.888234
10240	0.307200	0.317827	0.887575	0.887647	0.887418	0.888031
10752	0.300100	0.311810	0.890908	0.890827	0.890747	0.891181
11264	0.303400	0.311010	0.889871	0.887939	0.888309	0.887929
11776	0.300500	0.309282	0.891041	0.889819	0.890077	0.889860
12288	0.303600	0.326918	0.891272	0.891250	0.890942	0.891689
12800	0.300300	0.301688	0.894516	0.894619	0.894481	0.894940
13312	0.302200	0.302173	0.896441	0.896527	0.896462	0.896769
13824	0.299800	0.293489	0.895047	0.895172	0.895084	0.895448
14336	0.294600	0.297645	0.895865	0.896012	0.895886	0.896261
14848	0.296700	0.300751	0.895277	0.895401	0.895304	0.895651
15360	0.293100	0.293049	0.896855	0.896705	0.896757	0.896871
15872	0.293600	0.294201	0.895933	0.895557	0.895624	0.895651
16384	0.290100	0.289367	0.897847	0.897889	0.897840	0.898090
16896	0.293600	0.283990	0.898889	0.898724	0.898789	0.898903
17408	0.285800	0.308257	0.898250	0.898102	0.898162	0.898293
17920	0.252400	0.327164	0.898860	0.898807	0.898831	0.899004
18432	0.219500	0.315286	0.898877	0.898835	0.898831	0.899004
18944	0.217900	0.312738	0.898857	0.898958	0.898886	0.899207
19456	0.186400	0.320669	0.899252	0.899166	0.899194	0.899411
19968	0.199000	0.316840	0.900458	0.900455	0.900426	0.900630

Model Card Authors

Erfan Moosavi Monazzah