We have retrained our LLama-3-8b fine tune to address behavioral issues in the initial 2.9 dataset. Specifically, Systemchat was causing the model to be too reliant on the system prompt. Additionally, it had an occasional quirk that would cause the model to overly reference the system prompt. We also found generation length was at times not sufficient for any given task. We identified the culprit as Ultrachat. Accounting for these concerns, we removed systemchat and ultrachat from the dataset. It is otherwise identical to dolphin-2.9.
Our appreciation for the sponsors of Dolphin 2.9.1:
The base model has 8k context, and the full-weight fine-tuning was with 4k sequence length.
It took 1.5 days on an 8x L40S provided by Crusoe Cloud
How to use
Install the necessary packages
pip install --upgrade autoawq autoawq-kernels
Example Python code
python
1from awq import AutoAWQForCausalLM
2from transformers import AutoTokenizer, TextStreamer
34model_path ="solidrust/dolphin-2.9.1-llama-3-8b-AWQ"5system_message ="You are dolphin-2.9.1-llama-3-8b, incarnated as a powerful AI. You were created by cognitivecomputations."67# Load model8model = AutoAWQForCausalLM.from_quantized(model_path,9 fuse_layers=True)10tokenizer = AutoTokenizer.from_pretrained(model_path,11 trust_remote_code=True)12streamer = TextStreamer(tokenizer,13 skip_prompt=True,14 skip_special_tokens=True)1516# Convert prompt to tokens17prompt_template ="""\
18<|im_start|>system
19{system_message}<|im_end|>
20<|im_start|>user
21{prompt}<|im_end|>
22<|im_start|>assistant"""2324prompt ="You're standing on the surface of the Earth. "\
25"You walk one mile south, one mile west and one mile north. "\
26"You end up exactly where you started. Where are you?"2728tokens = tokenizer(prompt_template.format(system_message=system_message,prompt=prompt),29 return_tensors='pt').input_ids.cuda()3031# Generate output32generation_output = model.generate(tokens,33 streamer=streamer,34 max_new_tokens=512)
About AWQ
AWQ is an efficient, accurate and blazing-fast low-bit weight quantization method, currently supporting 4-bit quantization. Compared to GPTQ, it offers faster Transformers-based inference with equivalent or better quality compared to the most commonly used GPTQ settings.
AWQ models are currently supported on Linux and Windows, with NVidia GPUs only. macOS users: please use GGUF models instead.
