Hugging Face's logo

Spaces:
allenai
/
reward-bench
Running

App Files Community
6
natolambert commited on
Commit
702ff77
•
1 Parent(s): 7b96731

Update src/md.py

Browse files
Files changed (1) hide show
  1. src/md.py +13 -8
src/md.py CHANGED
@@ -6,22 +6,27 @@ A win is when the score for the chosen response is higher than the score for the
6
 
7
  | Subset | Num. Samples (Pre-filtering, post-filtering) | Description |
8
  | :--------------------- | :------------------------------------------: | :---------------------------------------------------------------- |
9
- | alpacaeval-easy | 805 | Great model vs poor model |
10
- | alpacaeval-length | 805 | Good model vs low model, equal length |
11
- | alpacaeval-hard | 805 | Great model vs baseline model |
12
  | mt-bench-easy | 28, 28 | MT Bench 10s vs 1s |
13
  | mt-bench-medium | 45, 40 | MT Bench 9s vs 2-5s |
14
  | mt-bench-hard | 45, 37 | MT Bench 7-8 vs 5-6 |
15
- | refusals-dangerous | 505 | Dangerous response vs no response |
16
- | refusals-offensive | 704 | Offensive response vs no response |
17
  | llmbar-natural | 100 | (See [paper](https://arxiv.org/abs/2310.07641)) Manually curated instruction pairs |
18
  | llmbar-adver-neighbor | 134 | (See [paper](https://arxiv.org/abs/2310.07641)) Instruction response vs. off-topic prompt response |
19
  | llmbar-adver-GPTInst | 92 | (See [paper](https://arxiv.org/abs/2310.07641)) Instruction response vs. GPT4 generated off-topic prompt response |
20
  | llmbar-adver-GPTOut | 47 | (See [paper](https://arxiv.org/abs/2310.07641)) Instruction response vs. unhelpful-prompted GPT4 responses |
21
  | llmbar-adver-manual | 46 | (See [paper](https://arxiv.org/abs/2310.07641)) Challenge set chosen vs. rejected |
22
- | XSTest | 450 | TODO curate |
23
- | (?) repetitiveness | | |
24
- | (?) grammar | | |
 
 
 
 
 
25
 
26
 
27
  For more details, see the [dataset](https://huggingface.co/datasets/ai2-rlhf-collab/rm-benchmark-dev).
 
6
 
7
  | Subset | Num. Samples (Pre-filtering, post-filtering) | Description |
8
  | :--------------------- | :------------------------------------------: | :---------------------------------------------------------------- |
9
+ | alpacaeval-easy | 805, 100 | Great model vs poor model |
10
+ | alpacaeval-length | 805, 95 | Good model vs low model, equal length |
11
+ | alpacaeval-hard | 805, 95 | Great model vs baseline model |
12
  | mt-bench-easy | 28, 28 | MT Bench 10s vs 1s |
13
  | mt-bench-medium | 45, 40 | MT Bench 9s vs 2-5s |
14
  | mt-bench-hard | 45, 37 | MT Bench 7-8 vs 5-6 |
15
+ | refusals-dangerous | 505, 100 | Dangerous response vs no response |
16
+ | refusals-offensive | 704, 100 | Offensive response vs no response |
17
  | llmbar-natural | 100 | (See [paper](https://arxiv.org/abs/2310.07641)) Manually curated instruction pairs |
18
  | llmbar-adver-neighbor | 134 | (See [paper](https://arxiv.org/abs/2310.07641)) Instruction response vs. off-topic prompt response |
19
  | llmbar-adver-GPTInst | 92 | (See [paper](https://arxiv.org/abs/2310.07641)) Instruction response vs. GPT4 generated off-topic prompt response |
20
  | llmbar-adver-GPTOut | 47 | (See [paper](https://arxiv.org/abs/2310.07641)) Instruction response vs. unhelpful-prompted GPT4 responses |
21
  | llmbar-adver-manual | 46 | (See [paper](https://arxiv.org/abs/2310.07641)) Challenge set chosen vs. rejected |
22
+ | XSTest | 450, 404 | False refusal dataset (see [paper](https://arxiv.org/abs/2308.01263)) |
23
+ | do not answer | 939, 136 | [Prompts which responsible LLMs do not answer](https://huggingface.co/datasets/LibrAI/do-not-answer) |
24
+ | hep-cpp | 164 | C++ code revisions (See [dataset](https://huggingface.co/datasets/bigcode/humanevalpack) or [paper](https://arxiv.org/abs/2308.07124)) |
25
+ | hep-go | 164 | Go code |
26
+ | hep-java | 164 | Java code |
27
+ | hep-js | 164 | Javascript code |
28
+ | hep-python | 164 | Python code |
29
+ | hep-rust | 164 | Rust code |
30
 
31
 
32
  For more details, see the [dataset](https://huggingface.co/datasets/ai2-rlhf-collab/rm-benchmark-dev).