# Graphcore/gptj-mnli

This model is the fine-tuned version of EleutherAI/gpt-j-6B on the GLUE MNLI dataset. MNLI dataset consists of pairs of sentences, a premise and a hypothesis. The task is to predict the relation between the premise and the hypothesis, which can be:

• entailment: hypothesis follows from the premise,
• contradiction: hypothesis contradicts the premise,
• neutral: hypothesis and premise are unrelated.

We finetune the model as a Causal Language Model (CLM): given a sequence of tokens, the task is to predict the next token. To achieve this, we create a stylised prompt string, following the approach of T5 paper.

mnli hypothesis: {hypothesis} premise: {premise} target: {class_label} <|endoftext|>


For example:

mnli hypothesis: Your contributions were of no help with our students' education. premise: Your contribution helped make it possible for us to provide our students with a quality education. target: contradiction <|endoftext|>


## Model description

GPT-J 6B is a transformer model trained using Ben Wang's Mesh Transformer JAX. "GPT-J" refers to the class of model, while "6B" represents the number of trainable parameters.

The model consists of 28 layers with a model dimension of 4096, and a feedforward dimension of 16384. The model dimension is split into 16 heads, each with a dimension of 256. Rotary Position Embedding (RoPE) is applied to 64 dimensions of each head. The model is trained with a tokenization vocabulary of 50257, using the same set of BPEs as GPT-2/GPT-3.

EleutherAI/gpt-j-6B, our starting point for finetuning, is trained on the Pile, a large-scale curated dataset created by EleutherAI.

## Fine-tuning and validation data

Fine tuning is done using the train split of the GLUE MNLI dataset and the performance is measured using the validation_mismatched split.

validation_mismatched means validation examples are not derived from the same sources as those in the training set and therefore not closely resembling any of the examples seen at training time.

Data splits for the mnli dataset are the following

train validation_matched validation_mismatched
392702 9815 9832

## Fine-tuning procedure

Fine tuned on a Graphcore IPU-POD64 using popxl.

Prompt sentences are tokenized and packed together to form 1024 token sequences, following HF packing algorithm. No padding is used. The packing process works in groups of 1000 examples and discards any remainder from each group that isn't a whole sequence. For the 392,702 training examples this gives a total of 17,762 sequences per epoch.

Since the model is trained to predict the next token, labels are simply the input sequence shifted by one token. Given the training format, no extra care is needed to account for different sequences: the model does not need to know which sentence a token belongs to.

### Hyperparameters:

• optimiser: AdamW (beta1: 0.9, beta2: 0.999, eps: 1e-6, weight decay: 0.0, learning rate: 5e-6)
• learning rate schedule: warmup schedule (min: 1e-7, max: 5e-6, warmup proportion: 0.005995)
• batch size: 128
• training steps: 300. Each epoch consists of ceil(17,762/128) steps, hence 300 steps are approximately 2 epochs.

## Performance

The resulting model matches SOTA performance with 82.5% accuracy.

Total number of examples                 9832
Number with badly formed result          0
Number with incorrect result             1725
Number with correct result               8107
[82.5%]

example 0 = {'prompt_text': "mnli hypothesis: Your contributions were of no help with our students' education. premise: Your contribution helped make it possible for us to provide our students with a quality education. target:", 'class_label': 'contradiction'}

First 10 generated_text and expected class_label results:
2: 'entailment'                             entailment
4: 'entailment'                             entailment
5: 'entailment'                             entailment
8: 'entailment'                             neutral


## How to use

The model can be easily loaded using AutoModelForCausalLM. You can use the pipeline API for text generation.

from transformers import pipeline, AutoModelForCausalLM, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained('EleutherAI/gpt-j-6B')
generator =  pipeline('text-generation', model=hf_model, tokenizer=tokenizer)
prompt = "mnli hypothesis: Your contributions were of no help with our students' education." \
"premise: Your contribution helped make it possible for us to provide our students with a quality education. target:"
out = generator(prompt, return_full_text=False, max_new_tokens=5, top_k=1)


You can create prompt-like inputs starting from GLUE MNLI dataset using functions provided in the data_utils.py script.

from datasets import load_dataset
from data_utils import form_text, split_text

dataset = dataset.map(
form_text, remove_columns=['hypothesis', 'premise','label', 'idx'])
# dataset[0] {'text': "mnli hypothesis: Your contributions were of no help with our students' education. premise: Your contribution helped make it possible for us to provide our students with a quality education. target: contradiction<|endoftext|>"}
dataset = dataset.map(split_text, remove_columns=['text'])
# dataset[0] {'prompt_text': "mnli hypothesis: Your contributions were of no help with our students' education. premise: Your contribution helped make it possible for us to provide our students with a quality education. target:",