| ## Training details | |
| - Dataset used: Explanation style datasets from psmathur/WizardLM_Orca and Dahoas/cot_gsm8k | |
| - Techniques: fp16 bit precision training + LoRA + DeepSpeed | |
| - Machine: V100 (16GB) * 2 | |
| ## Inference | |
| ```python | |
| from peft import PeftModel | |
| from huggingface_hub import hf_hub_download | |
| from transformers import GPTNeoForCausalLM, AutoTokenizer | |
| import json | |
| model_name = "shahules786/Redpajama-3B-orcastyle" | |
| config = hf_hub_download(repo_id=model_name, filename="adapter_config.json", local_dir=".") | |
| config = json.load(open("adapter_config.json")) | |
| base_model = config["base_model_name_or_path"] | |
| tokenizer = AutoTokenizer.from_pretrained(model_name) | |
| model = GPTNeoForCausalLM.from_pretrained(base_model) | |
| model.resize_token_embeddings(len(self.tokenizer)) | |
| model = PeftModel.from_pretrained(model, model_name).eval() | |
| tokenizer.padding_side = "left" | |
| inputs = tokenizer("This is a sample run", return_tensors="pt") | |
| model.generate(**inputs) | |
| ``` | |
| Checkout training and inference code [here](https://github.com/explodinggradients/Funtuner/tree/main/funtuner) | |