Original model description:

license: mit base_model: microsoft/phi-2 tags: - generated_from_trainer model-index: - name: phi-2-pl-v_0_1 results: []

phi-2-pl-v_0_1

This model is based on microsoft/phi-2. It was trained from scratch on the 20231201 Polish Wikipedia dump.

Model description

The model was trained for a context length of 2048 tokens.

Intended uses & limitations

The model is intended for research purposes only. It may generate fictitious, incorrect, unethical, or biased texts. At its current state, it is not suitable for production purposes.

Example:

tokenizer = AutoTokenizer.from_pretrained(
    model_name, trust_remote_code=True, use_fast=True
)
model = AutoModelForCausalLM.from_pretrained(
    model_name, vocab_size=len(tokenizer), attn_implementation="flash_attention_2",
    trust_remote_code=True, torch_dtype=torch.bfloat16, device_map="auto"
)
model.eval()

generation_config = GenerationConfig.from_pretrained(
    model_name, do_sample=False, repetition_penalty=1.5,
    min_new_tokens=1, max_new_tokens=128
)

test_input = tokenizer("Wrocław to polski miasto. Wrocław jest ", return_tensors='pt').to(torch.device('cuda'))
test_output = model.generate(**test_input, generation_config=generation_config)
test_preds = tokenizer.batch_decode(sequences=test_output, skip_special_tokens=True, clean_up_tokenization_spaces=True)
print(test_preds)

Training and evaluation data

The 20231201 Polish Wikipedia dump.

Training procedure

Training environment

GPU: 1 x A100X (80GB)

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 0.0002
num_devices: 1
train_batch_size: 8
gradient_accumulation_steps: 1
optimizer: Adam with betas=(0.9,0.98) and epsilon=1e-07
lr_scheduler_type: cosine
lr_scheduler_warmup_ratio: 0.1
num_epochs: 1
precision: bf16
seed: 42

Training results

runtime: 1mo 3d 9h 40m 16s
train_loss: 2.983

Framework versions

Transformers 4.37.1
Pytorch 2.1.2
Datasets 2.16.1
Tokenizers 0.15.1