Commit
•
fea86db
1
Parent(s):
27ce9e6
Update README.md
Browse files
README.md
CHANGED
|
@@ -8,4 +8,41 @@ license: llama3
|
|
| 8 |
|
| 9 |
The extrapolated (ExPO) model based on [`RLHFlow/LLaMA3-iterative-DPO-final`](https://huggingface.co/RLHFlow/LLaMA3-iterative-DPO-final) and [`RLHFlow/LLaMA3-SFT`](https://huggingface.co/RLHFlow/LLaMA3-SFT), as in the "[Weak-to-Strong Extrapolation Expedites Alignment](https://arxiv.org/abs/2404.16792)" paper.
|
| 10 |
|
| 11 |
-
Specifically, we obtain this model by extrapolating **(alpha = 0.3)** from the weights of the SFT and DPO/RLHF checkpoints, achieving superior alignment with human preference.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 8 |
|
| 9 |
The extrapolated (ExPO) model based on [`RLHFlow/LLaMA3-iterative-DPO-final`](https://huggingface.co/RLHFlow/LLaMA3-iterative-DPO-final) and [`RLHFlow/LLaMA3-SFT`](https://huggingface.co/RLHFlow/LLaMA3-SFT), as in the "[Weak-to-Strong Extrapolation Expedites Alignment](https://arxiv.org/abs/2404.16792)" paper.
|
| 10 |
|
| 11 |
+
Specifically, we obtain this model by extrapolating **(alpha = 0.3)** from the weights of the SFT and DPO/RLHF checkpoints, achieving superior alignment with human preference.
|
| 12 |
+
|
| 13 |
+
## Evaluation Results
|
| 14 |
+
|
| 15 |
+
Evaluation results on the **AlpacaEval 2.0** benchmark (you can find the evaluation outputs on the [official GitHub repo](https://github.com/chujiezheng/LLM-Extrapolation/tree/main/results_alpaca)):
|
| 16 |
+
|
| 17 |
+
| | Win Rate (Ori) | LC Win Rate (Ori) | Win Rate (+ ExPO) | LC Win Rate (+ ExPO) |
|
| 18 |
+
| ------------------------------------ | -------------- | ----------------- | ----------------- | -------------------- |
|
| 19 |
+
| `HuggingFaceH4/zephyr-7b-alpha` | 6.7% | 10.0% | **10.6%** | **13.6%** |
|
| 20 |
+
| `HuggingFaceH4/zephyr-7b-beta` | 10.2% | 13.2% | **11.1%** | **14.0%** |
|
| 21 |
+
| `berkeley-nest/Starling-LM-7B-alpha` | 15.0% | 18.3% | **18.2%** | **19.5%** |
|
| 22 |
+
| `Nexusflow/Starling-LM-7B-beta` | 26.6% | 25.8% | **29.6%** | **26.4%** |
|
| 23 |
+
| `snorkelai/Snorkel-Mistral-PairRM` | 24.7% | 24.0% | **28.8%** | **26.4%** |
|
| 24 |
+
| `RLHFlow/LLaMA3-iterative-DPO-final` | 29.2% | 36.0% | **32.7%** | **37.8%** |
|
| 25 |
+
| `internlm/internlm2-chat-1.8b` | 3.8% | 4.0% | **5.2%** | **4.3%** |
|
| 26 |
+
| `internlm/internlm2-chat-7b` | 20.5% | 18.3% | **28.1%** | **22.7%** |
|
| 27 |
+
| `internlm/internlm2-chat-20b` | 36.1% | 24.9% | **46.2%** | **27.2%** |
|
| 28 |
+
| `allenai/tulu-2-dpo-7b` | 8.5% | 10.2% | **11.5%** | **11.7%** |
|
| 29 |
+
| `allenai/tulu-2-dpo-13b` | 11.2% | 15.5% | **15.6%** | **17.6%** |
|
| 30 |
+
| `allenai/tulu-2-dpo-70b` | 15.4% | 21.2% | **23.0%** | **25.7%** |
|
| 31 |
+
|
| 32 |
+
Evaluation results on the **MT-Bench** benchmark (you can find the evaluation outputs on the [official GitHub repo](https://github.com/chujiezheng/LLM-Extrapolation/tree/main/results_mtbench)):
|
| 33 |
+
|
| 34 |
+
| | Original | + ExPO |
|
| 35 |
+
| ------------------------------------ | -------- | -------- |
|
| 36 |
+
| `HuggingFaceH4/zephyr-7b-alpha` | 6.85 | **6.87** |
|
| 37 |
+
| `HuggingFaceH4/zephyr-7b-beta` | 7.02 | **7.06** |
|
| 38 |
+
| `berkeley-nest/Starling-LM-7B-alpha` | 7.82 | **7.91** |
|
| 39 |
+
| `Nexusflow/Starling-LM-7B-beta` | 8.10 | **8.18** |
|
| 40 |
+
| `snorkelai/Snorkel-Mistral-PairRM` | 7.63 | **7.69** |
|
| 41 |
+
| `RLHFlow/LLaMA3-iterative-DPO-final` | 8.08 | **8.45** |
|
| 42 |
+
| `internlm/internlm2-chat-1.8b` | 5.17 | **5.26** |
|
| 43 |
+
| `internlm/internlm2-chat-7b` | 7.72 | **7.80** |
|
| 44 |
+
| `internlm/internlm2-chat-20b` | 8.13 | **8.26** |
|
| 45 |
+
| `allenai/tulu-2-dpo-7b` | 6.35 | **6.38** |
|
| 46 |
+
| `allenai/tulu-2-dpo-13b` | 7.00 | **7.26** |
|
| 47 |
+
| `allenai/tulu-2-dpo-70b` | 7.79 | **8.03** |
|
| 48 |
+
|