from transformers import AutoModelForSequenceClassification, AutoTokenizer
import torch

device = "cuda" if torch.cuda.is_available() else "cpu"
checkpoint = "nie3e/sentiment-polish-gpt2-small"
tokenizer = AutoTokenizer.from_pretrained(checkpoint)
model = AutoModelForSequenceClassification.from_pretrained(
    checkpoint
).to(device)

text = "Jak na cenę - super . Śniadanie nie do przejedzenia ."

input_ids = tokenizer(text, return_tensors="pt").to(device)
logits = model(**input_ids)["logits"].to("cpu")

percent = torch.sigmoid(logits).squeeze(dim=0)

id2class = model.config.id2label
print({id2class[i]: f"{(p*100):.2f}%" for i, p in enumerate(percent.tolist())})

{'NEUTRAL': '2.95%', 'NEGATIVE': '0.21%', 'POSITIVE': '100.00%', 'AMBIGUOUS': '20.56%'}

Transformers Pipeline

from transformers import pipeline

pipe = pipeline(
    "sentiment-analysis",
    "nie3e/sentiment-polish-gpt2-small"
)

result = pipe("Jak na cenę - super . Śniadanie nie do przejedzenia .")
print(result)

[{'label': 'POSITIVE', 'score': 0.9999890327453613}]

vLLM >= 0.9.2

from vllm import LLM
llm = LLM(
    "nie3e/sentiment-polish-gpt2-small",
    task="classify",
    enforce_eager=True
)

text = ["Jak na cenę - super . Śniadanie nie do przejedzenia ."]
outputs = llm.classify(text)

for output in outputs:
    print(output.outputs.probs)

vLLM OpenAI serving (recommended)

docker run --gpus 1 --ipc=host -p 8000:8000 vllm/vllm-openai:v0.9.2 --model nie3e/sentiment-polish-gpt2-small

using curl:

bash

curl -X 'POST' \
  'http://127.0.0.1:8000/classify' \
  -H 'Content-Type: application/json' \
  -d '{
  "model": "nie3e/sentiment-polish-gpt2-small",
  "input": ["Przestronny hotel , jasny , z dużymi oknami .", "Położony całkiem blisko centrum ."]
}'

result:

json

{
	"id": "classify-6619cecdb01a4bf8900df136a9b33b15",
	"object": "list",
	"created": 1749994841,
	"model": "nie3e/sentiment-polish-gpt2-small",
	"data": [
		{
			"index": 0,
			"label": "POSITIVE",
			"probs": [
				0.000006198883056640625,
				1.7881393432617188e-7,
				1.0,
				0.000007569789886474609
			],
			"num_classes": 4
		},
		{
			"index": 1,
			"label": "AMBIGUOUS",
			"probs": [
				0.00013005733489990234,
				0.004421234130859375,
				0.005367279052734375,
				0.990234375
			],
			"num_classes": 4
		}
	],
	"usage": {
		"prompt_tokens": 17,
		"total_tokens": 17,
		"completion_tokens": 0,
		"prompt_tokens_details": null
	}
}

using python:

import requests

response = requests.post(
    f"http://127.0.0.1:8000/classify",
    headers={"Content-Type": "application/json"},
    json={
        "model": "nie3e/sentiment-polish-gpt2-small",
        "input": [
            "Przestronny hotel , jasny , z dużymi oknami .",
            "Położony całkiem blisko centrum ."
        ]
    }
)

print(response.json())

{'id': 'classify-ac86189c5d0e41908584c3e88d356316',
 'object': 'list',
 'created': 1753209126,
 'model': 'nie3e/sentiment-polish-gpt2-small',
 'data': [{'index': 0,
   'label': 'POSITIVE',
   'probs': [6.277556167333387e-06,
    1.9292743047572003e-07,
    0.9999858140945435,
    7.631567314092536e-06],
   'num_classes': 4},
  {'index': 1,
   'label': 'AMBIGUOUS',
   'probs': [0.0001290593936573714,
    0.004407374653965235,
    0.005339722614735365,
    0.9901238083839417],
   'num_classes': 4}],
 'usage': {'prompt_tokens': 17,
  'total_tokens': 17,
  'completion_tokens': 0,
  'prompt_tokens_details': None}}

Training and evaluation data

Merged all rows from polemo2-official dataset.

Train/test split: 80%/20%

Datacollator:

from transformers import DataCollatorWithPadding
data_collator = DataCollatorWithPadding(
  tokenizer=tokenizer,
  padding="longest",
  max_length=128,
  pad_to_multiple_of=8
)

Training procedure

GPU: RTX 3090

Training time: 2:53:05

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 2e-05
train_batch_size: 4
eval_batch_size: 4
seed: 42
gradient_accumulation_steps: 8
total_train_batch_size: 32
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: linear
num_epochs: 10
mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Accuracy
0.4049	1.0	3284	0.3351	0.8792
0.1885	2.0	6568	0.2625	0.9218
0.1182	3.0	9852	0.2583	0.9419
0.0825	4.0	13136	0.2886	0.9482
0.0586	5.0	16420	0.3343	0.9538
0.034	6.0	19704	0.3734	0.9595
0.0288	7.0	22988	0.4125	0.9599
0.0185	8.0	26273	0.4262	0.9626
0.0069	9.0	29557	0.4529	0.9622
0.0059	10.0	32840	0.4659	0.9627

Evaluation

Evaluated on allegro/klej-polemo2-out test dataset.

from datasets import load_dataset
from evaluate import evaluator

data = load_dataset("allegro/klej-polemo2-out", split="test").shuffle(seed=42)
task_evaluator = evaluator("text-classification")

# fix labels
l = {
        "__label__meta_zero": 0,
        "__label__meta_minus_m": 1,
        "__label__meta_plus_m": 2,
        "__label__meta_amb": 3
    }
def fix_labels(examples):
    examples["target"] = l[examples["target"]]
    return examples
data = data.map(fix_labels)

eval_resutls = task_evaluator.compute(
    model_or_pipeline="nie3e/sentiment-polish-gpt2-small",
    data=data,
    label_mapping={"NEUTRAL": 0, "NEGATIVE": 1, "POSITIVE": 2, "AMBIGUOUS": 3},
    input_column="sentence",
    label_column="target"
)

print(eval_resutls)

json

{
    "accuracy": 0.9838056680161943,
    "total_time_in_seconds": 5.2441766999982065,
    "samples_per_second": 94.1997244296076,
    "latency_in_seconds": 0.010615742307688678
}

Framework versions

Transformers 4.36.2
Pytorch 2.1.2+cu121
Datasets 2.16.1
Tokenizers 0.15.0