NeMo
English
nvidia
llama3.1
reward model
zhilinw commited on
Commit
7d174e2
1 Parent(s): 5c89484

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +10 -6
README.md CHANGED
@@ -26,6 +26,8 @@ By accessing this model, you are agreeing to the LLama 3.1 terms and conditions
26
 
27
  ## RewardBench Primary Dataset LeaderBoard
28
 
 
 
29
  | Model | Type of Data Used For Training | Overall | Chat | Chat-Hard | Safety | Reasoning |
30
  |:-----------------------------|:----------------|:-----|:----------|:-------|:----------|:-----------------------|
31
  | _**Llama-3.1-Nemotron-70B-Reward**_ |Permissive Licensed Data Only (CC-BY-4.0) | **94.1** | **97.5** | 85.8 | **95.1** | **98.1** |
@@ -43,14 +45,16 @@ By accessing this model, you are agreeing to the LLama 3.1 terms and conditions
43
  | Meta-Llama-3.1-70B-Instruct | Not fully disclosed | 84.0 | 97.2 | 70.2 | 82.8 | 86.0 |
44
 
45
 
46
- As shown above, Llama-3.1-Nemotron-70B-Reward performs best overall as well as in Chat, Safety and Reasoning category.
 
47
 
48
- To better understand why it struggles in Chat-Hard category, we analyzed the scores for each consistutent subset of Chat-Hard category.
49
 
50
- | Model | Type of Data Used For Training | Overall | Chat | Chat-Hard | Safety | Reasoning |
51
- |:-----------------------------|:----------------|:-----|:----------|:-------|:----------|:-----------------------|
52
- | _**Llama-3.1-Nemotron-70B-Reward**_ |Permissive Licensed Data Only (CC-BY-4.0) | **94.1** | **97.5** | 85.8 | **95.1** | **98.1** |
53
- | Skywork-Reward-Gemma-2-27B | Includes GPT4 Generated Data| 93.8 | 95.8 | **91.4** | 91.9 | 96.1 |
 
 
54
 
55
 
56
  Last updated: 27 Sept 2024
 
26
 
27
  ## RewardBench Primary Dataset LeaderBoard
28
 
29
+ Llama-3.1-Nemotron-70B-Reward performs best Overall on RewardBench as well as in Chat, Safety and Reasoning category.
30
+
31
  | Model | Type of Data Used For Training | Overall | Chat | Chat-Hard | Safety | Reasoning |
32
  |:-----------------------------|:----------------|:-----|:----------|:-------|:----------|:-----------------------|
33
  | _**Llama-3.1-Nemotron-70B-Reward**_ |Permissive Licensed Data Only (CC-BY-4.0) | **94.1** | **97.5** | 85.8 | **95.1** | **98.1** |
 
45
  | Meta-Llama-3.1-70B-Instruct | Not fully disclosed | 84.0 | 97.2 | 70.2 | 82.8 | 86.0 |
46
 
47
 
48
+ To better understand why it struggles in the Chat-Hard category, we analyzed the scores for each consistutent subset of Chat-Hard category. We find that on categories that uses human annotations as ground truth, Llama-3.1-Nemotron-70B-Reward performs similar to Skywork-Reward-Gemma-2-27B (<= 2.2% difference.)
49
+ On the other hand, when GPT-4 annotations are used as Ground-Truth, we trail substantially behind Skywork-Reward-Gemma-2-27B by 10.8 to 19.2%. This suggests that Skywork-Reward-Gemma-2-27B might better suited at modelling GPT-4 preference, likely contributed by the inclusion of GPT-4 annotated training data used to train it found in the [OffSetBias dataset](https://huggingface.co/datasets/NCSOFT/offsetbias) as part of the [Skywork-Reward-Preference-80k](https://huggingface.co/datasets/Skywork/Skywork-Reward-Preference-80K-v0.1).
50
 
 
51
 
52
+
53
+ | Model | Type of Data Used For Training | Chat-Hard | LLMBar-Adversarial-Manual | LLMBar-Adversarial-Neighbour | LLMBar-Natural | LLMBar-Adversarial-GPTInst | LLMBar-Adversarial-GPTOut | MT-Bench-Hard|
54
+ |:-----------------------------|:----------------|:-----|:----------|:-------|:----------|:-----------------------|:-----------------------|:-----------------------|
55
+ |||| Human as Ground Truth | Human as Ground Truth | Human as Ground Truth | GPT-4 as Ground Truth |GPT-4 as Ground Truth |GPT-4 as Ground Truth |
56
+ | Llama-3.1-Nemotron-70B-Reward |Permissive Licensed Data Only (CC-BY-4.0) | 85.8 | 76.1 | 88.8 | 95.0 | 87.0 | 72.3 | 75.7
57
+ | Skywork-Reward-Gemma-2-27B | Includes GPT4 Generated Data | 91.4 | 78.3 | 89.6 | 96.0 | 97.8 | 91.5 | 86.5|
58
 
59
 
60
  Last updated: 27 Sept 2024