Update README.md
Browse files
README.md
CHANGED
@@ -6,28 +6,88 @@ tags:
|
|
6 |
- orpo
|
7 |
- generated_from_trainer
|
8 |
model-index:
|
9 |
-
- name:
|
10 |
results: []
|
11 |
---
|
12 |
|
13 |
<!-- This model card has been generated automatically according to the information the Trainer had access to. You
|
14 |
should probably proofread and complete it, then remove this comment. -->
|
15 |
|
16 |
-
#
|
17 |
|
18 |
-
|
19 |
|
20 |
-
## Model
|
21 |
|
22 |
-
|
23 |
|
24 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
25 |
|
26 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
27 |
|
28 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
29 |
|
30 |
-
More information needed
|
31 |
|
32 |
## Training procedure
|
33 |
|
|
|
6 |
- orpo
|
7 |
- generated_from_trainer
|
8 |
model-index:
|
9 |
+
- name: zephyr-orpo-141b-A35b-v0.1
|
10 |
results: []
|
11 |
---
|
12 |
|
13 |
<!-- This model card has been generated automatically according to the information the Trainer had access to. You
|
14 |
should probably proofread and complete it, then remove this comment. -->
|
15 |
|
16 |
+
# Model Card for Zephyr 141B-A35B
|
17 |
|
18 |
+
Zephyr is a series of language models that are trained to act as helpful assistants. Zephyr 141B-A35B is the latest model in the series, and is a fine-tuned version of [mistral-community/Mixtral-8x22B-v0.1](https://huggingface.co/mistral-community/Mixtral-8x22B-v0.1) that was trained using a novel alignment algorithm called alignment algorithm called [Odds Ratio Preference Optimization (ORPO)](https://huggingface.co/papers/2403.07691). ORPO does not require an SFT step to achieve high performance and is thus much more computationally efficient than methods like DPO and PPO. To train Zephyr-141B-A35B, we used the [`argilla/distilabel-capybara-dpo-7k-binarized`](https://huggingface.co/datasets/argilla/distilabel-capybara-dpo-7k-binarized) preference dataset, which consists of synthetic, high-quality, multi-turn preferences that have been scored via LLMs.
|
19 |
|
20 |
+
## Model Details
|
21 |
|
22 |
+
### Model Description
|
23 |
|
24 |
+
<!-- Provide a longer summary of what this model is. -->
|
25 |
+
|
26 |
+
- **Model type:** A Mixture of Experts (MoE) model with 141B total parameters and 35B active parameters. Fine-tuned on a mix of publicly available, synthetic datasets.
|
27 |
+
- **Language(s) (NLP):** Primarily English.
|
28 |
+
- **License:** Apache 2.0
|
29 |
+
- **Finetuned from model:** [mistral-community/Mixtral-8x22B-v0.1](https://huggingface.co/mistral-community/Mixtral-8x22B-v0.1)
|
30 |
+
|
31 |
+
### Model Sources
|
32 |
+
|
33 |
+
<!-- Provide the basic links for the model. -->
|
34 |
+
|
35 |
+
- **Repository:** https://github.com/huggingface/alignment-handbook
|
36 |
+
- **Dataset:** https://huggingface.co/datasets/argilla/distilabel-capybara-dpo-7k-binarized
|
37 |
|
38 |
+
## Performance
|
39 |
+
|
40 |
+
Zephyr 141B-A35B was trained to test the effectiveness of ORPO at scale and the underlying dataset contains a mix of general chat capabilities. It achieves strong performance on chat benchmarks like [MT Bench](https://huggingface.co/spaces/lmsys/mt-bench) and [IFEval](https://arxiv.org/abs/2311.07911). The scores reported below were obtained using the [LightEval](https://github.com/huggingface/lighteval) evaluation suite and each prompt has been formatted with the model's corresponding chat template to simulate real-world usage. This is why some scores may differ from those reported in technical reports or on the Open LLM Leaderboard.
|
41 |
+
|
42 |
+
| Model | MT Bench | IFEval | BBH | AGIEval |
|
43 |
+
|-----------------------------------------------------------------------------------------------------|---------:|-------:|------:|--------:|
|
44 |
+
| [zephyr-orpo-141b-A35b-v0.1](https://huggingface.co/HuggingFaceH4/zephyr-orpo-141b-A35b-v0.1) | 8.14 | 65.06 | 58.96 | 44.16 |
|
45 |
+
| [databricks/dbrx-instruct](https://huggingface.co/databricks/dbrx-instruct) | 8.26 | 52.13 | 48.50 | 41.16 |
|
46 |
+
| [mistralai/Mixtral-8x7B-Instruct-v0.1](https://huggingface.co/mistralai/Mixtral-8x7B-Instruct-v0.1) | 8.30 | 55.08 | 45.31 | 47.68 |
|
47 |
+
|
48 |
+
|
49 |
+
## Intended uses & limitations
|
50 |
|
51 |
+
The model was fine-tuned on a blend of chat, code, math, and reasoning data. Here's how you can run the model using the `pipeline()` function from 🤗 Transformers:
|
52 |
+
|
53 |
+
```python
|
54 |
+
# pip install 'transformers>=4.39.3'
|
55 |
+
# pip install accelerate
|
56 |
+
|
57 |
+
import torch
|
58 |
+
from transformers import pipeline
|
59 |
+
|
60 |
+
pipe = pipeline(
|
61 |
+
"text-generation",
|
62 |
+
model="orpo-explorers/zephyr-orpo-141b-A35b-v0.1",
|
63 |
+
device_map="auto",
|
64 |
+
torch_dtype=torch.bfloat16,
|
65 |
+
)
|
66 |
+
messages = [
|
67 |
+
{
|
68 |
+
"role": "system",
|
69 |
+
"content": "You are Zephyr, a helpful assistant.",
|
70 |
+
},
|
71 |
+
{"role": "user", "content": "What can you tell me about black holes?"},
|
72 |
+
]
|
73 |
+
outputs = pipe(
|
74 |
+
messages,
|
75 |
+
max_new_tokens=512,
|
76 |
+
do_sample=True,
|
77 |
+
temperature=0.7,
|
78 |
+
top_k=50,
|
79 |
+
top_p=0.95,
|
80 |
+
)
|
81 |
+
print(outputs[0]["generated_text"][-1]["content"])
|
82 |
+
```
|
83 |
+
|
84 |
+
## Bias, Risks, and Limitations
|
85 |
+
|
86 |
+
<!-- This section is meant to convey both technical and sociotechnical limitations. -->
|
87 |
+
|
88 |
+
Zephyr 141B-A35B has not been aligned to human preferences for safety within the RLHF phase or deployed with in-the-loop filtering of responses like ChatGPT, so the model can produce problematic outputs (especially when prompted to do so).
|
89 |
+
It is also unknown what the size and composition of the corpus was used to train the base model (`mistral-community/Mixtral-8x22B-v0.1`), however it is likely to have included a mix of Web data and technical sources like books and code. See the [Falcon 180B model card](https://huggingface.co/tiiuae/falcon-180B#training-data) for an example of this.
|
90 |
|
|
|
91 |
|
92 |
## Training procedure
|
93 |
|