Update README.md

README.md

@@ -1,3 +1,72 @@
 ---
 license: apache-2.0
----
+datasets:
+- mmosiolek/pl_alpaca_data_cleaned
+language:
+- pl
+tags:
+- alpaca
+- llama
+- self-instruct
+- casual language model
+- llm
+- gpt
+- chat-gpt
+---
+# Polpaca: The Alpaca Speaks Polish
+
+Dataset for the project: https://huggingface.co/datasets/mmosiolek/pl_alpaca_data_cleaned
+
+[LLaMA](https://ai.facebook.com/blog/large-language-model-llama-meta-ai/) is a state-of-the-art, foundational, open-source large language model designed to help engineers and researchers advance their work in NLP. 
+For example, Stanford researchers have fine-tuned LLaMA to construct an alternative to the famous ChatGPT - a model called [Alpaca](https://crfm.stanford.edu/2023/03/13/alpaca.html).
+Unfortunately, [LLaMA](https://ai.facebook.com/blog/large-language-model-llama-meta-ai/) was trained on a dataset consisting mainly of English texts, with only 4.5% of the data relating to other languages.
+In addition, the [Alpaca](https://crfm.stanford.edu/2023/03/13/alpaca.html) instruction training dataset consists only of examples of English instructions.
+So [Alpaca](https://crfm.stanford.edu/2023/03/13/alpaca.html) simply doesn't work for the other languages.
+
+This repo makes [Alpaca-Lora-7B](https://huggingface.co/tloen/alpaca-lora-7b) speak Polish. 
+
+### Usage
+
+```python
+from transformers import LlamaTokenizer, LlamaForCausalLM
+from peft import PeftModel
+import bitsandbytes as bnb
+
+base = "decapoda-research/llama-7b-hf"
+finetuned = "mmosiolek/polpaca-lora-7b"
+
+tokenizer = LlamaTokenizer.from_pretrained(base)
+tokenizer.pad_token_id = 0
+tokenizer.padding_side = "left"
+
+model = LlamaForCausalLM.from_pretrained(base)
+model = PeftModel.from_pretrained(model, finetuned).to("cuda")
+```
+
+For output generation use the following code:
+
+
+```python
+from transformers import GenerationConfig
+
+config = GenerationConfig(
+  temperature=0.1,
+  top_p=0.75,
+  top_k=40,
+  num_beams=4,
+  max_new_tokens=128,
+)
+
+def run(instruction, model, tokenizer):
+    encodings = tokenizer(instruction, padding=True, return_tensors="pt").to('cuda')
+    generated_ids = model.generate(
+        **encodings,
+        generation_config=GENERATION_CONFIG,
+    )
+    decoded = tokenizer.batch_decode(generated_ids)
+    del encodings, generated_ids
+    torch.cuda.empty_cache()
+    return decoded[0].split("\n")[-1]
+```
+
+Example input/output