98 lines
3.7 KiB
Markdown
98 lines
3.7 KiB
Markdown
---
|
|
tags:
|
|
- generated_from_trainer
|
|
- conversational
|
|
- polish
|
|
license: mit
|
|
language:
|
|
- pl
|
|
datasets:
|
|
- eryk-mazus/polka-dpo-v1
|
|
pipeline_tag: text-generation
|
|
inference: false
|
|
---
|
|
|
|
|
|
|
|
# Polka-1.1B-Chat
|
|
|
|
`eryk-mazus/polka-1.1b-chat` **is the first polish model trained to act as a helpful, conversational assistant that can be run locally.**
|
|
|
|
The model is based on [TinyLlama-1.1B](https://huggingface.co/TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T) 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)](https://arxiv.org/abs/2305.18290) performed on top of it.**
|
|
|
|
Context size: 4,096 tokens
|
|
|
|
In addition, we're releasing:
|
|
* [polka-1.1b](https://huggingface.co/eryk-mazus/polka-1.1b) - our base model with an extended tokenizer and additional pre-training on Polish corpus sampled using [DSIR](https://github.com/p-lambda/dsir)
|
|
* [polka-pretrain-en-pl-v1](https://huggingface.co/datasets/eryk-mazus/polka-pretrain-en-pl-v1) - the pre-training dataset
|
|
* [polka-dpo-v1](https://huggingface.co/datasets/eryk-mazus/polka-dpo-v1) - dataset of DPO pairs
|
|
* [polka-1.1b-chat-gguf](https://huggingface.co/eryk-mazus/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](https://github.com/vllm-project/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](https://huggingface.co/docs/transformers/chat_templating), which means you can format messages using the `tokenizer.apply_chat_template()` method, as demonstrated in the example above.
|