The model is based on TinyLlama-1.1B with the custom, extended tokenizer for more efficient Polish text generation, that was additionally pretrained on 5.7 billion tokens. It was then fine-tuned on around 60k synthetically generated and machine-translated multi-turn conversations with the Direct Preference Optimization (DPO) performed on top of it.

Context size: 4,096 tokens

In addition, we're releasing:

polka-1.1b - our base model with an extended tokenizer and additional pre-training on Polish corpus sampled using DSIR
polka-pretrain-en-pl-v1 - the pre-training dataset
polka-dpo-v1 - dataset of DPO pairs
polka-1.1b-chat-gguf - GGUF files for the chat model

Usage

Sample code:

python

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM, TextStreamer

model_name = "eryk-mazus/polka-1.1b-chat"

tokenizer = AutoTokenizer.from_pretrained(model_name, use_fast=True)
tokenizer.pad_token = tokenizer.eos_token

model = AutoModelForCausalLM.from_pretrained(
    model_name, 
    torch_dtype=torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float16,
    device_map="auto"
)
streamer = TextStreamer(tokenizer, skip_prompt=True)

# You are a helpful assistant.
system_prompt = "Jesteś pomocnym asystentem."
chat = [{"role": "system", "content": system_prompt}]

# Compose a short song on programming.
user_input = "Napisz krótką piosenkę o programowaniu."
chat.append({"role": "user", "content": user_input})

# Generate - add_generation_prompt to make sure it continues as assistant
inputs = tokenizer.apply_chat_template(chat, add_generation_prompt=True, return_tensors="pt")
# For multi-GPU, find the device of the first parameter of the model
first_param_device = next(model.parameters()).device
inputs = inputs.to(first_param_device)

with torch.no_grad():
    outputs = model.generate(
        inputs,
        pad_token_id=tokenizer.eos_token_id,
        max_new_tokens=512,
        temperature=0.2,
        repetition_penalty=1.15,
        top_p=0.95,
        do_sample=True,
        streamer=streamer,
    )

# Add just the new tokens to our chat
new_tokens = outputs[0, inputs.size(1):]
response = tokenizer.decode(new_tokens, skip_special_tokens=True)
chat.append({"role": "assistant", "content": response})

The model works seamlessly with vLLM as well.

Prompt format

This model uses ChatML as the prompt format:

<|im_start|>system
Jesteś pomocnym asystentem.
<|im_start|>user
Jakie jest dzienne zapotrzebowanie kaloryczne dorosłej osoby?<|im_end|>
<|im_start|>assistant
Dla dorosłych osób zaleca się spożywanie około 2000-3000 kcal dziennie, aby utrzymać optymalne zdrowie i dobre samopoczucie.<|im_end|>

This prompt is available as a chat template, which means you can format messages using the tokenizer.apply_chat_template() method, as demonstrated in the example above.