vikarti-anatra's picture
initial commit
df12d9a
|
raw
history blame
4.27 kB
metadata
language:
  - en
  - zh
license: mit
datasets:
  - wenbopan/Chinese-dpo-pairs
  - Intel/orca_dpo_pairs
  - argilla/ultrafeedback-binarized-preferences-cleaned
  - jondurbin/truthy-dpo-v0.1
pipeline_tag: text-generation

Vikarti notes

EXL2 6bpw version of https://huggingface.co/wenbopan/Faro-Yi-9B-DPO

This is my first "quant"

Thanks to https://new.reddit.com/user/Downtown-Case-1755/ for making it known in https://new.reddit.com/r/LocalLLaMA/comments/1cziy0m/what_is_sota_for_a_mega_context_100k_novel/ that long-context models could be used by GPU-poor people who doesn't have 3090/4090.

Faro-Yi-9B-DPO

This is the DPO version of wenbopan/Faro-Yi-9B. Compared to Faro-Yi-9B and Yi-9B-200K, the DPO model excels at many tasks, surpassing the original Yi-9B-200K by a large margin. On the Open LLM Leaderboard, it ranks #2 among all 9B models, #1 among all Yi-9B variants.

Metric MMLU GSM8K hellaswag truthfulqa ai2_arc winogrande CMMLU
Yi-9B-200K 65.73 50.49 56.72 33.80 69.25 71.67 71.97
Faro-Yi-9B 68.80 63.08 57.28 40.86 72.58 71.11 73.28
Faro-Yi-9B-DPO 69.98 66.11 59.04 48.01 75.68 73.40 75.23

Faro-Yi-9B-DPO's responses are also favored by GPT-4 Judge in MT-Bench

image/png

How to Use

Faro-Yi-9B-DPO uses the chatml template and performs well in both short and long contexts. For longer inputs under 24GB of VRAM, I recommend to use vLLM to have a max prompt of 32K. Setting kv_cache_dtype="fp8_e5m2" allows for 48K input length. 4bit-AWQ quantization on top of that can boost input length to 160K, albeit with some performance impact. Adjust max_model_len arg in vLLM or config.json to avoid OOM.

import io
import requests
from PyPDF2 import PdfReader
from vllm import LLM, SamplingParams

llm = LLM(model="wenbopan/Faro-Yi-9B-DPO", kv_cache_dtype="fp8_e5m2", max_model_len=100000)

pdf_data = io.BytesIO(requests.get("https://arxiv.org/pdf/2303.08774.pdf").content)
document = "".join(page.extract_text() for page in PdfReader(pdf_data).pages) # 100 pages

question = f"{document}\n\nAccording to the paper, what is the parameter count of GPT-4?"
messages = [ {"role": "user", "content": question} ] # 83K tokens
prompt = llm.get_tokenizer().apply_chat_template(messages, add_generation_prompt=True, tokenize=False)
output = llm.generate(prompt, SamplingParams(temperature=0.8, max_tokens=500))
print(output[0].outputs[0].text)
# Yi-9B-200K:      175B. GPT-4 has 175B \nparameters. How many models were combined to create GPT-4? Answer: 6. ...
# Faro-Yi-9B: GPT-4 does not have a publicly disclosed parameter count due to the competitive landscape and safety implications of large-scale models like GPT-4. ...
Or With Transformers
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained('wenbopan/Faro-Yi-9B-DPO', device_map="cuda")
tokenizer = AutoTokenizer.from_pretrained('wenbopan/Faro-Yi-9B-DPO')
messages = [
    {"role": "system", "content": "You are a helpful assistant. Always answer with a short response."},
    {"role": "user", "content": "Tell me what is Pythagorean theorem like you are a pirate."}
]

input_ids = tokenizer.apply_chat_template(messages, tokenize=True, add_generation_prompt=True, return_tensors="pt").to(model.device)
generated_ids = model.generate(input_ids, max_new_tokens=512, temperature=0.5)
response = tokenizer.decode(generated_ids[0], skip_special_tokens=True) # Aye, matey! The Pythagorean theorem is a nautical rule that helps us find the length of the third side of a triangle. ...