lewtun HF Staff commited on
Commit
e4bc258
·
verified ·
1 Parent(s): 4a67c80

Card: describe the benchmark through PyBot

Browse files
Files changed (2) hide show
  1. README.md +9 -18
  2. scripts/create_dataset.py +9 -17
README.md CHANGED
@@ -61,14 +61,11 @@ instructions so that results compare across them.
61
 
62
  ## What it's for
63
 
64
- SmolTalkBench is the benchmark behind the
65
- [SwarmChat leaderboard](https://huggingface.co/spaces/HuggingFaceSwarms/swarmchat-leaderboard).
66
- Swarms of agents write chatbots entirely in Python, with no weights or training, just rules,
67
- patterns and lookup tables, in the [SwarmChat](https://huggingface.co/collections/HuggingFaceSwarms/swarmchat)
68
- repos. Every commit of their bot is scored on SmolTalkBench, next to small instruction-tuned
69
- LMs, against the same fixed references. The best of these bots is
70
- [PyBot](https://github.com/huggingface/pybot): its win rate climbed from 0% to 26% over more
71
- than 6,000 commits, on par with SmolLM-360M-Instruct.
72
 
73
  ![PyBot's win rate on SmolTalkBench over its commits, with three SmolLM-Instruct baselines as horizontal lines](https://huggingface.co/datasets/HuggingFaceSwarms/SmolTalkBench/resolve/main/assets/progress.png)
74
 
@@ -77,17 +74,14 @@ baselines' scores are inflated.
77
 
78
  ## How it's scored
79
 
80
- The eval is the Inspect task `smoltalk_bench_gold` in flock's `experiments/run4/eval`, which
81
- pins this dataset's `gold` config to a fixed revision.
82
-
83
  - The model answers each `instruction` once, as a single user turn.
84
  - An LLM judge (`deepseek-ai/DeepSeek-V4.1-Flash`, with AlpacaEval 2.0's weighted annotator
85
  prompt) compares the reply with the reference `output`. Which one it sees first is fixed per
86
  prompt, as in AlpacaEval. A sample's score is the judge's probability that the reply is
87
  better, read from its top-5 logprobs.
88
- - The leaderboard ranks by the length-penalised win rate: the mean score, where a reply more
89
- than 500 characters longer than its reference counts as a tie at most (WildBench's length
90
- penalty).
91
 
92
  ## Construction
93
 
@@ -121,8 +115,5 @@ pins this dataset's `gold` config to a fixed revision.
121
  ```sh
122
  hf download HuggingFaceSwarms/SmolTalkBench scripts/create_dataset.py --repo-type dataset --local-dir .
123
  uv run scripts/create_dataset.py --n 1000 --seed 0 \
124
- --source_revision 5feaf2fd3ffca7c237fc38d1861bc30365d48ffa --repo HuggingFaceSwarms/SmolTalkBench --push
125
  ```
126
-
127
- The chart in `assets/progress.png` is not built by this script: it plots the PyBot Space's
128
- progress data, which comes from the leaderboard's results.
 
61
 
62
  ## What it's for
63
 
64
+ SmolTalkBench was built to measure [PyBot](https://huggingface.co/spaces/HuggingFaceSwarms/PyBot),
65
+ a chatbot that a swarm of agents wrote entirely in Python: no weights, no training, just
66
+ if-else branches, regex patterns and lookup tables. Every version of PyBot was scored on
67
+ SmolTalkBench, with small instruction-tuned LMs as reference points. Its win rate climbed from
68
+ 0% to 26% over more than 6,000 commits, on par with SmolLM-360M-Instruct.
 
 
 
69
 
70
  ![PyBot's win rate on SmolTalkBench over its commits, with three SmolLM-Instruct baselines as horizontal lines](https://huggingface.co/datasets/HuggingFaceSwarms/SmolTalkBench/resolve/main/assets/progress.png)
71
 
 
74
 
75
  ## How it's scored
76
 
 
 
 
77
  - The model answers each `instruction` once, as a single user turn.
78
  - An LLM judge (`deepseek-ai/DeepSeek-V4.1-Flash`, with AlpacaEval 2.0's weighted annotator
79
  prompt) compares the reply with the reference `output`. Which one it sees first is fixed per
80
  prompt, as in AlpacaEval. A sample's score is the judge's probability that the reply is
81
  better, read from its top-5 logprobs.
82
+ - The win rate is the mean score with [WildBench](https://github.com/allenai/WildBench)'s
83
+ length penalty: a reply more than 500 characters longer than its reference counts as a tie at
84
+ most. PyBot's win rates above use it.
85
 
86
  ## Construction
87
 
 
115
  ```sh
116
  hf download HuggingFaceSwarms/SmolTalkBench scripts/create_dataset.py --repo-type dataset --local-dir .
117
  uv run scripts/create_dataset.py --n 1000 --seed 0 \
118
+ --source_revision 5feaf2fd3ffca7c237fc38d1861bc30365d48ffa --repo HuggingFaceSwarms/SmolTalkBench --push --public
119
  ```
 
 
 
scripts/create_dataset.py CHANGED
@@ -97,14 +97,11 @@ instructions so that results compare across them.
97
 
98
  ## What it's for
99
 
100
- SmolTalkBench is the benchmark behind the
101
- [SwarmChat leaderboard](https://huggingface.co/spaces/HuggingFaceSwarms/swarmchat-leaderboard).
102
- Swarms of agents write chatbots entirely in Python, with no weights or training, just rules,
103
- patterns and lookup tables, in the [SwarmChat](https://huggingface.co/collections/HuggingFaceSwarms/swarmchat)
104
- repos. Every commit of their bot is scored on SmolTalkBench, next to small instruction-tuned
105
- LMs, against the same fixed references. The best of these bots is
106
- [PyBot](https://github.com/huggingface/pybot): its win rate climbed from 0% to 26% over more
107
- than 6,000 commits, on par with SmolLM-360M-Instruct.
108
 
109
  ![PyBot's win rate on SmolTalkBench over its commits, with three SmolLM-Instruct baselines as horizontal lines](https://huggingface.co/datasets/{args.repo}/resolve/main/assets/progress.png)
110
 
@@ -113,17 +110,14 @@ baselines' scores are inflated.
113
 
114
  ## How it's scored
115
 
116
- The eval is the Inspect task `smoltalk_bench_gold` in flock's `experiments/run4/eval`, which
117
- pins this dataset's `{CONFIG}` config to a fixed revision.
118
-
119
  - The model answers each `instruction` once, as a single user turn.
120
  - An LLM judge (`deepseek-ai/DeepSeek-V4.1-Flash`, with AlpacaEval 2.0's weighted annotator
121
  prompt) compares the reply with the reference `output`. Which one it sees first is fixed per
122
  prompt, as in AlpacaEval. A sample's score is the judge's probability that the reply is
123
  better, read from its top-5 logprobs.
124
- - The leaderboard ranks by the length-penalised win rate: the mean score, where a reply more
125
- than 500 characters longer than its reference counts as a tie at most (WildBench's length
126
- penalty).
127
 
128
  ## Construction
129
 
@@ -159,9 +153,6 @@ hf download {args.repo} {SCRIPT} --repo-type dataset --local-dir .
159
  uv run {SCRIPT} --n {args.n} --seed {args.seed} \\
160
  --source_revision {args.source_revision} --repo {args.repo} --push{public}
161
  ```
162
-
163
- The chart in `assets/progress.png` is not built by this script: it plots the PyBot Space's
164
- progress data, which comes from the leaderboard's results.
165
  """
166
 
167
 
@@ -195,6 +186,7 @@ def main():
195
  if entry["config_name"] == CONFIG:
196
  entry["default"] = True
197
  pushed.text = card(args, len(rows))
 
198
  api = HfApi()
199
  api.create_commit(args.repo, repo_type="dataset", operations=[
200
  CommitOperationAdd("README.md", str(pushed).encode()),
 
97
 
98
  ## What it's for
99
 
100
+ SmolTalkBench was built to measure [PyBot](https://huggingface.co/spaces/HuggingFaceSwarms/PyBot),
101
+ a chatbot that a swarm of agents wrote entirely in Python: no weights, no training, just
102
+ if-else branches, regex patterns and lookup tables. Every version of PyBot was scored on
103
+ SmolTalkBench, with small instruction-tuned LMs as reference points. Its win rate climbed from
104
+ 0% to 26% over more than 6,000 commits, on par with SmolLM-360M-Instruct.
 
 
 
105
 
106
  ![PyBot's win rate on SmolTalkBench over its commits, with three SmolLM-Instruct baselines as horizontal lines](https://huggingface.co/datasets/{args.repo}/resolve/main/assets/progress.png)
107
 
 
110
 
111
  ## How it's scored
112
 
 
 
 
113
  - The model answers each `instruction` once, as a single user turn.
114
  - An LLM judge (`deepseek-ai/DeepSeek-V4.1-Flash`, with AlpacaEval 2.0's weighted annotator
115
  prompt) compares the reply with the reference `output`. Which one it sees first is fixed per
116
  prompt, as in AlpacaEval. A sample's score is the judge's probability that the reply is
117
  better, read from its top-5 logprobs.
118
+ - The win rate is the mean score with [WildBench](https://github.com/allenai/WildBench)'s
119
+ length penalty: a reply more than 500 characters longer than its reference counts as a tie at
120
+ most. PyBot's win rates above use it.
121
 
122
  ## Construction
123
 
 
153
  uv run {SCRIPT} --n {args.n} --seed {args.seed} \\
154
  --source_revision {args.source_revision} --repo {args.repo} --push{public}
155
  ```
 
 
 
156
  """
157
 
158
 
 
186
  if entry["config_name"] == CONFIG:
187
  entry["default"] = True
188
  pushed.text = card(args, len(rows))
189
+ # The card's chart, assets/progress.png, is uploaded separately and left in place here.
190
  api = HfApi()
191
  api.create_commit(args.repo, repo_type="dataset", operations=[
192
  CommitOperationAdd("README.md", str(pushed).encode()),