HuggingFaceSwarms/SmolTalkBench

subset	description	num_samples
gold (default)	Single-turn everyday chat prompts from smoltalk's everyday-conversations, sampled in proportion to their parent topic, with the source conversations' replies by meta-llama/Meta-Llama-3.1-70B-Instruct as references	1000

1000 single-turn everyday chat prompts with reference replies, for AlpacaEval-style pairwise win rates.

from datasets import load_dataset

ds = load_dataset("HuggingFaceSwarms/SmolTalkBench", "gold", split="test")

Variants with other reference models go in their own configs, keeping the same ids and instructions so that results compare across them.

What it's for

SmolTalkBench was built to measure PyBot, a chatbot that a swarm of agents wrote entirely in Python: no weights, no training, just if-else branches, regex patterns and lookup tables. Every version of PyBot was scored on SmolTalkBench, with small instruction-tuned LMs as reference points. Its win rate climbed from 0% to 26% over more than 6,000 commits, on par with SmolLM-360M-Instruct.

SmolLM and SmolLM2 Instruct models were trained on the source conversations, so the SmolLM baselines' scores are inflated.

How it's scored

• The model answers each instruction once, as a single user turn.
• An LLM judge (deepseek-ai/DeepSeek-V4.1-Flash, with AlpacaEval 2.0's weighted annotator prompt) compares the reply with the reference output. Which one it sees first is fixed per prompt, as in AlpacaEval. A sample's score is the judge's probability that the reply is better, read from its top-5 logprobs.
• The win rate is the mean score with WildBench's length penalty: a reply more than 500 characters longer than its reference counts as a tie at most. PyBot's win rates above use it.

Construction

• Source: the everyday-conversations subset of HuggingFaceTB/smoltalk at revision 5feaf2fd3ffc, train and test splits. Each conversation was written whole by meta-llama/Meta-Llama-3.1-70B-Instruct.
• Every source conversation opens with a bare greeting ("Hi", "Hello", ...) answered by "Hello! How can I help you today?". The greeting is dropped: instruction is the second user turn and output the assistant's reply to it.
• Prompts are deduplicated (case and whitespace insensitive), leaving 2198 rows.
• 1000 rows are sampled with seed 0, in proportion to each parent topic (topic without its last level), so every subcategory keeps its share.

Columns

column	description
id	<source split>-<source index>
instruction	the prompt (second user turn)
output	reference reply by generator
generator, dataset	reference model and source, as in AlpacaEval's reference outputs
topic	the source full_topic
category	first level of topic
group	everyday (three-level topics) or school (two-level grade-school knowledge topics)
source_split, source_index	the row in the source subset

Reproducing

scripts/create_dataset.py builds the gold config and pushes it with this card and itself:

hf download HuggingFaceSwarms/SmolTalkBench scripts/create_dataset.py --repo-type dataset --local-dir .
uv run scripts/create_dataset.py --n 1000 --seed 0 \
  --source_revision 5feaf2fd3ffca7c237fc38d1861bc30365d48ffa --repo HuggingFaceSwarms/SmolTalkBench --push --public