Commit
•
4331d2e
1
Parent(s):
871c7d0
Update README.md
Browse files
README.md
CHANGED
|
@@ -1,71 +1,81 @@
|
|
| 1 |
-
---
|
| 2 |
-
language:
|
| 3 |
-
- en
|
| 4 |
-
- zh
|
| 5 |
-
license: mit
|
| 6 |
-
datasets:
|
| 7 |
-
- wenbopan/Chinese-dpo-pairs
|
| 8 |
-
- Intel/orca_dpo_pairs
|
| 9 |
-
- argilla/ultrafeedback-binarized-preferences-cleaned
|
| 10 |
-
- jondurbin/truthy-dpo-v0.1
|
| 11 |
-
pipeline_tag: text-generation
|
| 12 |
-
---
|
| 13 |
-
|
| 14 |
-
|
| 15 |
-
|
| 16 |
-
|
| 17 |
-
|
| 18 |
-
|
| 19 |
-
|
| 20 |
-
|
| 21 |
-
|
| 22 |
-
|
| 23 |
-
|
| 24 |
-
Faro-Yi-9B-DPO
|
| 25 |
-
|
| 26 |
-
|
| 27 |
-
|
| 28 |
-
|
| 29 |
-
|
| 30 |
-
|
| 31 |
-
|
| 32 |
-
|
| 33 |
-
|
| 34 |
-
|
| 35 |
-
|
| 36 |
-
|
| 37 |
-
|
| 38 |
-
|
| 39 |
-
|
| 40 |
-
|
| 41 |
-
|
| 42 |
-
|
| 43 |
-
|
| 44 |
-
|
| 45 |
-
|
| 46 |
-
|
| 47 |
-
|
| 48 |
-
|
| 49 |
-
|
| 50 |
-
|
| 51 |
-
|
| 52 |
-
|
| 53 |
-
|
| 54 |
-
|
| 55 |
-
|
| 56 |
-
|
| 57 |
-
|
| 58 |
-
|
| 59 |
-
|
| 60 |
-
|
| 61 |
-
|
| 62 |
-
|
| 63 |
-
|
| 64 |
-
|
| 65 |
-
|
| 66 |
-
|
| 67 |
-
|
| 68 |
-
|
| 69 |
-
|
| 70 |
-
|
| 71 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
language:
|
| 3 |
+
- en
|
| 4 |
+
- zh
|
| 5 |
+
license: mit
|
| 6 |
+
datasets:
|
| 7 |
+
- wenbopan/Chinese-dpo-pairs
|
| 8 |
+
- Intel/orca_dpo_pairs
|
| 9 |
+
- argilla/ultrafeedback-binarized-preferences-cleaned
|
| 10 |
+
- jondurbin/truthy-dpo-v0.1
|
| 11 |
+
pipeline_tag: text-generation
|
| 12 |
+
---
|
| 13 |
+
|
| 14 |
+
|
| 15 |
+
# Vikarti notes
|
| 16 |
+
EXL2 8bpw version of https://huggingface.co/wenbopan/Faro-Yi-9B-DPO
|
| 17 |
+
|
| 18 |
+
This is my first "quant"
|
| 19 |
+
|
| 20 |
+
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.
|
| 21 |
+
|
| 22 |
+
|
| 23 |
+
|
| 24 |
+
# Faro-Yi-9B-DPO
|
| 25 |
+
|
| 26 |
+
This is the DPO version of [wenbopan/Faro-Yi-9B](https://huggingface.co/wenbopan/Faro-Yi-9B). Compared to Faro-Yi-9B and [Yi-9B-200K](https://huggingface.co/01-ai/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](https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard), it ranks **#2** among all 9B models, **#1** among all Yi-9B variants.
|
| 27 |
+
|
| 28 |
+
| **Metric** | **MMLU** | **GSM8K** | **hellaswag** | **truthfulqa** | **ai2_arc** | **winogrande** | **CMMLU** |
|
| 29 |
+
| ----------------------- | --------- | --------- | ------------- | -------------- | ----------- | -------------- | --------- |
|
| 30 |
+
| **Yi-9B-200K** | 65.73 | 50.49 | 56.72 | 33.80 | 69.25 | 71.67 | 71.97 |
|
| 31 |
+
| **Faro-Yi-9B** | 68.80 | 63.08 | 57.28 | 40.86 | 72.58 | 71.11 | 73.28 |
|
| 32 |
+
| **Faro-Yi-9B-DPO** | **69.98** | **66.11** | **59.04** | **48.01** | **75.68** | **73.40** | **75.23** |
|
| 33 |
+
|
| 34 |
+
Faro-Yi-9B-DPO's responses are also favored by GPT-4 Judge in MT-Bench
|
| 35 |
+
|
| 36 |
+
|
| 37 |
+
|
| 38 |
+
## How to Use
|
| 39 |
+
|
| 40 |
+
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.
|
| 41 |
+
|
| 42 |
+
|
| 43 |
+
```python
|
| 44 |
+
import io
|
| 45 |
+
import requests
|
| 46 |
+
from PyPDF2 import PdfReader
|
| 47 |
+
from vllm import LLM, SamplingParams
|
| 48 |
+
|
| 49 |
+
llm = LLM(model="wenbopan/Faro-Yi-9B-DPO", kv_cache_dtype="fp8_e5m2", max_model_len=100000)
|
| 50 |
+
|
| 51 |
+
pdf_data = io.BytesIO(requests.get("https://arxiv.org/pdf/2303.08774.pdf").content)
|
| 52 |
+
document = "".join(page.extract_text() for page in PdfReader(pdf_data).pages) # 100 pages
|
| 53 |
+
|
| 54 |
+
question = f"{document}\n\nAccording to the paper, what is the parameter count of GPT-4?"
|
| 55 |
+
messages = [ {"role": "user", "content": question} ] # 83K tokens
|
| 56 |
+
prompt = llm.get_tokenizer().apply_chat_template(messages, add_generation_prompt=True, tokenize=False)
|
| 57 |
+
output = llm.generate(prompt, SamplingParams(temperature=0.8, max_tokens=500))
|
| 58 |
+
print(output[0].outputs[0].text)
|
| 59 |
+
# Yi-9B-200K: 175B. GPT-4 has 175B \nparameters. How many models were combined to create GPT-4? Answer: 6. ...
|
| 60 |
+
# 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. ...
|
| 61 |
+
```
|
| 62 |
+
|
| 63 |
+
|
| 64 |
+
<details> <summary>Or With Transformers</summary>
|
| 65 |
+
|
| 66 |
+
```python
|
| 67 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer
|
| 68 |
+
|
| 69 |
+
model = AutoModelForCausalLM.from_pretrained('wenbopan/Faro-Yi-9B-DPO', device_map="cuda")
|
| 70 |
+
tokenizer = AutoTokenizer.from_pretrained('wenbopan/Faro-Yi-9B-DPO')
|
| 71 |
+
messages = [
|
| 72 |
+
{"role": "system", "content": "You are a helpful assistant. Always answer with a short response."},
|
| 73 |
+
{"role": "user", "content": "Tell me what is Pythagorean theorem like you are a pirate."}
|
| 74 |
+
]
|
| 75 |
+
|
| 76 |
+
input_ids = tokenizer.apply_chat_template(messages, tokenize=True, add_generation_prompt=True, return_tensors="pt").to(model.device)
|
| 77 |
+
generated_ids = model.generate(input_ids, max_new_tokens=512, temperature=0.5)
|
| 78 |
+
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. ...
|
| 79 |
+
```
|
| 80 |
+
|
| 81 |
+
</details>
|