diff --git a/CONFIG.md b/CONFIG.md
index ab4e093..5ef1bef 100644
--- a/CONFIG.md
+++ b/CONFIG.md
@@ -26,6 +26,7 @@ For example, if you want to override default parameters for the built-in PubMedQ
 # Configure the model and sample limit for the built-in PubMedQA benchmark.
 
 [benchmarks.pubmedqa]
+dataset = "hf://quantiles/PubMedQA"
 samples = 50
 model = "openai:gpt-5.6"
 ```
@@ -41,6 +42,7 @@ Built-in benchmarks run natively inside the CLI, without any custom code. Below
 | Field         | Type            | Required | Description                                                      |
 | ------------- | --------------- | -------- | ---------------------------------------------------------------- |
 | `type`        | string          | no       | Defaults to `"builtin"`. May be omitted for built-in benchmarks. |
+| `dataset`     | string          | yes      | Dataset source. Hugging Face datasets must use `hf://...` or `huggingface://...`. |
 | `samples`     | integer         | no       | Number of dataset rows to evaluate.                              |
 | `model`       | string or table | no       | Model sampler. See [model format](#model-format).                |
 | `max_workers` | integer         | no       | Maximum concurrent workers.                                      |
@@ -48,10 +50,24 @@ Built-in benchmarks run natively inside the CLI, without any custom code. Below
 If none of these fields are customized, the built-in benchmark uses the following defaults:
 
 - `type`: `builtin`
+- `dataset`: The benchmark's default Hugging Face dataset, used when no config section is provided.
 - `samples`: All samples available in the benchmark's dataset, in order
 - `model`: The "demo" model, which outputs random values
 - `max_workers`: The default parallelism provided by the Rust [Tokio runtime](https://tokio.rs/)
 
+#### `dataset` naming
+
+The `dataset` field accepts a Hugging Face dataset source prefixed with `hf://`
+or `huggingface://`. For example:
+
+```toml
+dataset = "hf://quantiles/PubMedQA"
+```
+
+The CLI resolves `hf://quantiles/PubMedQA` to the Hugging Face dataset ID
+`quantiles/PubMedQA` before using the normal Hugging Face download path.
+Other URI-style prefixes are rejected.
+
 #### `model` naming
 
 The `model` field described above accepts a provider-prefixed string, for example:
@@ -156,6 +172,7 @@ When you run `qt resume <run_id>`, the CLI looks up the stored eval name and inp
 
 ```toml
 [benchmarks.pubmedqa]
+dataset = "hf://quantiles/PubMedQA"
 model = "openai:gpt-5.6"
 ```
 
@@ -163,6 +180,7 @@ model = "openai:gpt-5.6"
 
 ```toml
 [benchmarks.simpleqa-verified]
+dataset = "hf://quantiles/simpleqa-verified"
 samples = 10
 ```
 
diff --git a/README.md b/README.md
index f0c781f..2da6ce6 100644
--- a/README.md
+++ b/README.md
@@ -79,6 +79,7 @@ For **built-in benchmarks**, configure settings like `samples`, `model`, and `ma
 
 ```toml
 [benchmarks.pubmedqa]
+dataset = "hf://quantiles/PubMedQA"
 samples = 50
 model = "openai:gpt-5.6"
 max_workers = 100
diff --git a/cli/README.md b/cli/README.md
index 8c75c28..934f8cc 100644
--- a/cli/README.md
+++ b/cli/README.md
@@ -60,10 +60,11 @@ You can customize how the CLI executes built-in benchmarks and custom evaluation
 
 ### Built-in benchmarks
 
-For built-in benchmarks, configure settings like `samples`, `model`, and `max_workers`:
+For built-in benchmarks, configure settings like `dataset`, `samples`, `model`, and `max_workers`:
 
 ```toml
 [benchmarks.pubmedqa]
+dataset = "hf://quantiles/PubMedQA"
 samples = 50
 model = "openai:gpt-5.6"
 max_workers = 100
diff --git a/cli/examples/configs/anthropic/quantiles.toml b/cli/examples/configs/anthropic/quantiles.toml
index 7405860..05f6e8a 100644
--- a/cli/examples/configs/anthropic/quantiles.toml
+++ b/cli/examples/configs/anthropic/quantiles.toml
@@ -5,6 +5,8 @@
 # variable. To get an API key, go to https://platform.claude.com/settings/workspaces/default/keys
 
 [benchmarks.pubmedqa]
+dataset = "hf://quantiles/PubMedQA"
+
 # The full expert-labeled PubMedQA benchmark has 1000 samples. Since Anthropic charges for
 # usage, we are limiting the number of samples to 50 in this demo, to keep costs lower.
 samples = 50
diff --git a/cli/examples/configs/both_error/.quantiles.toml b/cli/examples/configs/both_error/.quantiles.toml
index e2a96a1..040121c 100644
--- a/cli/examples/configs/both_error/.quantiles.toml
+++ b/cli/examples/configs/both_error/.quantiles.toml
@@ -1,8 +1,11 @@
 [benchmarks.pubmedqa]
+dataset = "hf://quantiles/PubMedQA"
 samples = 10
 
 [benchmarks.simpleqa-verified]
+dataset = "hf://quantiles/simpleqa-verified"
 samples = 100
 
 [benchmarks.financebench]
+dataset = "hf://quantiles/financebench"
 samples = 50
diff --git a/cli/examples/configs/cloudflare/quantiles.toml b/cli/examples/configs/cloudflare/quantiles.toml
index 29bf1cd..ae4f45c 100644
--- a/cli/examples/configs/cloudflare/quantiles.toml
+++ b/cli/examples/configs/cloudflare/quantiles.toml
@@ -13,6 +13,8 @@
 # https://dash.cloudflare.com/<account_id>/ai/ai-gateway
 
 [benchmarks.pubmedqa]
+dataset = "hf://quantiles/PubMedQA"
+
 # The full expert-labeled PubMedQA benchmark has 1000 samples. Since Cloudflare charges for
 # usage, we are limiting the number of samples to 50 in this demo, to keep costs lower.
 samples = 50
diff --git a/cli/examples/configs/gemini/quantiles.toml b/cli/examples/configs/gemini/quantiles.toml
index cfc348c..6c5c893 100644
--- a/cli/examples/configs/gemini/quantiles.toml
+++ b/cli/examples/configs/gemini/quantiles.toml
@@ -7,6 +7,8 @@
 # https://aistudio.google.com/prompts/new_chat
 
 [benchmarks.pubmedqa]
+dataset = "hf://quantiles/PubMedQA"
+
 # The full expert-labeled PubMedQA benchmark has 1000 samples. Since Google charges for
 # usage, we are limiting the number of samples to 50 in this demo, to keep costs lower.
 samples = 50
diff --git a/cli/examples/configs/openai/quantiles.toml b/cli/examples/configs/openai/quantiles.toml
index db46a7a..4e31d8d 100644
--- a/cli/examples/configs/openai/quantiles.toml
+++ b/cli/examples/configs/openai/quantiles.toml
@@ -8,6 +8,8 @@
 # https://platform.openai.com/home
 
 [benchmarks.pubmedqa]
+dataset = "hf://quantiles/PubMedQA"
+
 # The full expert-labeled PubMedQA benchmark has 1000 samples. Since OpenAI charges for
 # usage, we are limiting the number of samples to 50 in this demo, to keep costs lower.
 samples = 50
diff --git a/cli/examples/configs/simple/quantiles.toml b/cli/examples/configs/simple/quantiles.toml
index 06b287e..9f2998f 100644
--- a/cli/examples/configs/simple/quantiles.toml
+++ b/cli/examples/configs/simple/quantiles.toml
@@ -8,7 +8,9 @@
 # number of dataset rows processed.
 
 [benchmarks.pubmedqa]
+dataset = "hf://quantiles/PubMedQA"
 samples = 10
 
 [benchmarks.simpleqa-verified]
+dataset = "hf://quantiles/simpleqa-verified"
 samples = 100
diff --git a/cli/src/builtins/common.rs b/cli/src/builtins/common.rs
index 53ea370..6ed37fa 100644
--- a/cli/src/builtins/common.rs
+++ b/cli/src/builtins/common.rs
@@ -18,6 +18,9 @@ pub(crate) struct BuiltinConfig {
     /// Number of dataset rows to evaluate. If omitted, the entire dataset is used.
     #[serde(default)]
     pub(crate) limit: Option<usize>,
+    /// Dataset source to evaluate. Builtins currently support Hugging Face via `hf://...`.
+    #[serde(default)]
+    pub(crate) dataset: Option<String>,
     /// Which model sampler to use. If omitted, the builtin chooses a sensible default.
     #[serde(default)]
     pub(crate) model: Option<Sampler>,
diff --git a/cli/src/builtins/input.rs b/cli/src/builtins/input.rs
index 7d443b5..437a124 100644
--- a/cli/src/builtins/input.rs
+++ b/cli/src/builtins/input.rs
@@ -5,6 +5,7 @@ use crate::llm::Sampler;
 /// Normalized run input schema for all builtins.
 #[derive(Serialize)]
 pub(crate) struct BuiltinRunInput {
+    pub(crate) dataset: String,
     pub(crate) model: String,
     pub(crate) num_samples: usize,
     #[serde(skip_serializing_if = "Option::is_none")]
@@ -15,11 +16,13 @@ pub(crate) struct BuiltinRunInput {
 pub(crate) async fn set_builtin_run_input(
     db: &sea_orm::DatabaseConnection,
     run_id: i64,
+    dataset: &str,
     model: Option<&Sampler>,
     num_samples: usize,
     max_workers: Option<usize>,
 ) -> anyhow::Result<()> {
     let input = serde_json::to_string(&BuiltinRunInput {
+        dataset: dataset.to_owned(),
         model: builtin_model_name(model),
         num_samples,
         max_workers,
diff --git a/cli/src/builtins/pubmedqa/mod.rs b/cli/src/builtins/pubmedqa/mod.rs
index 7ea162c..d6e6cc5 100644
--- a/cli/src/builtins/pubmedqa/mod.rs
+++ b/cli/src/builtins/pubmedqa/mod.rs
@@ -7,7 +7,7 @@ use crate::builtins::dataset_runner::DatasetRunner;
 use crate::builtins::input::set_builtin_run_input;
 use crate::builtins::output::set_builtin_run_output;
 use crate::builtins::{BuiltinContext, BuiltinWorkflow};
-use crate::dataset::DatasetManager;
+use crate::dataset::{DatasetManager, resolve_hf_dataset_source};
 use crate::llm::LLMSampler;
 use crate::llm::random_label::RandomLabelSampler;
 
@@ -22,6 +22,8 @@ mod eval;
 /// `PubMedQA` builtin using the quantiles/PubMedQA dataset.
 pub struct PubmedqaBuiltin;
 
+const DEFAULT_DATASET_SOURCE: &str = "hf://quantiles/PubMedQA";
+
 #[expect(clippy::too_many_lines)]
 #[async_trait::async_trait]
 impl BuiltinWorkflow for PubmedqaBuiltin {
@@ -49,7 +51,12 @@ impl BuiltinWorkflow for PubmedqaBuiltin {
         };
 
         let manager = DatasetManager::new()?;
-        let dataset_id = "quantiles/PubMedQA";
+        let dataset_source = config
+            .base
+            .dataset
+            .as_deref()
+            .unwrap_or(DEFAULT_DATASET_SOURCE);
+        let dataset_id = resolve_hf_dataset_source(dataset_source)?;
         let info = manager
             .init(dataset_id, Some("pqa_labeled"), Some("train"), None)
             .await?;
@@ -62,6 +69,7 @@ impl BuiltinWorkflow for PubmedqaBuiltin {
         set_builtin_run_input(
             ctx.db,
             ctx.run_id,
+            dataset_source,
             config.base.model.as_ref(),
             limit,
             config.base.max_workers,
diff --git a/cli/src/builtins/similarity.rs b/cli/src/builtins/similarity.rs
index 051a260..0fd9083 100644
--- a/cli/src/builtins/similarity.rs
+++ b/cli/src/builtins/similarity.rs
@@ -9,7 +9,7 @@ use crate::builtins::dataset_runner::DatasetRunner;
 use crate::builtins::input::set_builtin_run_input;
 use crate::builtins::output::set_builtin_run_output;
 use crate::builtins::{BuiltinContext, BuiltinWorkflow};
-use crate::dataset::DatasetManager;
+use crate::dataset::{DatasetManager, resolve_hf_dataset_source};
 use crate::llm::LLMSampler;
 use crate::llm::random::RandomSampler;
 use crate::similarity::{
@@ -45,7 +45,7 @@ struct RowOutput {
 #[derive(Clone, Copy)]
 pub struct SimilarityBenchmark {
     name: &'static str,
-    dataset_id: &'static str,
+    dataset_source: &'static str,
     input_field: &'static str,
     target_field: &'static str,
 }
@@ -53,7 +53,7 @@ pub struct SimilarityBenchmark {
 /// `simpleqa-verified` builtin.
 pub const SIMPLEQA: SimilarityBenchmark = SimilarityBenchmark {
     name: "simpleqa-verified",
-    dataset_id: "quantiles/simpleqa-verified",
+    dataset_source: "hf://quantiles/simpleqa-verified",
     input_field: "problem",
     target_field: "answer",
 };
@@ -61,7 +61,7 @@ pub const SIMPLEQA: SimilarityBenchmark = SimilarityBenchmark {
 /// `financebench` builtin.
 pub const FINANCEBENCH: SimilarityBenchmark = SimilarityBenchmark {
     name: "financebench",
-    dataset_id: "quantiles/financebench",
+    dataset_source: "hf://quantiles/financebench",
     input_field: "question",
     target_field: "answer",
 };
@@ -96,7 +96,13 @@ impl BuiltinWorkflow for SimilarityBenchmark {
         };
 
         let manager = DatasetManager::new()?;
-        let info = manager.init(self.dataset_id, None, None, None).await?;
+        let dataset_source = config
+            .base
+            .dataset
+            .as_deref()
+            .unwrap_or(self.dataset_source);
+        let dataset_id = resolve_hf_dataset_source(dataset_source)?;
+        let info = manager.init(dataset_id, None, None, None).await?;
 
         let total = info
             .total_rows
@@ -109,6 +115,7 @@ impl BuiltinWorkflow for SimilarityBenchmark {
         set_builtin_run_input(
             db,
             run_id,
+            dataset_source,
             config.base.model.as_ref(),
             limit,
             config.base.max_workers,
@@ -120,7 +127,7 @@ impl BuiltinWorkflow for SimilarityBenchmark {
         let metric = &metric;
         let max_workers = config.base.max_workers.unwrap_or_else(get_max_workers);
 
-        let scores = DatasetRunner::new(&manager, self.dataset_id, &info, limit)
+        let scores = DatasetRunner::new(&manager, dataset_id, &info, limit)
             .desc(self.name)
             .set_quiet(ctx.quiet)
             .for_each_concurrent(max_workers, |i, row| {
diff --git a/cli/src/commands/resume.rs b/cli/src/commands/resume.rs
index 91b61af..86e68ed 100644
--- a/cli/src/commands/resume.rs
+++ b/cli/src/commands/resume.rs
@@ -132,6 +132,7 @@ mod tests {
         let bench = qt::config::BenchmarkConfig::Builtin(qt::config::BuiltinBenchmarkConfig {
             type_: "builtin".to_owned(),
             samples: None,
+            dataset: "hf://quantiles/PubMedQA".to_owned(),
             model: None,
             max_workers: None,
         });
@@ -146,6 +147,7 @@ mod tests {
         let bench = qt::config::BenchmarkConfig::Builtin(qt::config::BuiltinBenchmarkConfig {
             type_: "builtin".to_owned(),
             samples: Some(10),
+            dataset: "hf://quantiles/PubMedQA".to_owned(),
             model: None,
             max_workers: None,
         });
diff --git a/cli/src/commands/run.rs b/cli/src/commands/run.rs
index 18fcb06..ab536d8 100644
--- a/cli/src/commands/run.rs
+++ b/cli/src/commands/run.rs
@@ -103,12 +103,9 @@ fn assemble_builtin_input(
     }
 
     if let Some(bench) = bench {
-        if bench.samples.is_none() && bench.model.is_none() && bench.max_workers.is_none() {
-            return (None, Vec::new());
-        }
-
         let input = BuiltinConfigInput {
             limit: bench.samples,
+            dataset: bench.dataset.clone(),
             model: bench.model.clone(),
             max_workers: bench.max_workers,
         };
@@ -382,6 +379,7 @@ struct BuiltinRunJsonOutput {
 struct BuiltinConfigInput {
     #[serde(skip_serializing_if = "Option::is_none")]
     limit: Option<usize>,
+    dataset: String,
     #[serde(skip_serializing_if = "Option::is_none")]
     model: Option<qt::llm::Sampler>,
     #[serde(skip_serializing_if = "Option::is_none")]
@@ -623,6 +621,7 @@ mod tests {
         let bench = qt::config::BuiltinBenchmarkConfig {
             type_: "builtin".to_owned(),
             samples: Some(10),
+            dataset: "hf://quantiles/PubMedQA".to_owned(),
             model: None,
             max_workers: None,
         };
@@ -637,28 +636,32 @@ mod tests {
         let bench = qt::config::BuiltinBenchmarkConfig {
             type_: "builtin".to_owned(),
             samples: Some(5),
+            dataset: "hf://quantiles/PubMedQA".to_owned(),
             model: Some(qt::llm::Sampler::Random {}),
             max_workers: Some(8),
         };
         let (input, _) = super::assemble_builtin_input(Some(&bench), None);
         let parsed: serde_json::Value = serde_json::from_str(&input.unwrap()).unwrap();
         assert_eq!(parsed["limit"], 5);
+        assert_eq!(parsed["dataset"], "hf://quantiles/PubMedQA");
         assert_eq!(parsed["model"], "random");
         assert_eq!(parsed["max_workers"], 8);
     }
 
-    /// When the builtin config section exists but has no runtime-relevant fields, the
-    /// input should be `None` rather than an empty JSON object.
+    /// When the builtin config section only has the required dataset, the input should
+    /// still carry that dataset source into builtin execution.
     #[test]
-    fn assemble_builtin_input_none_when_no_config_fields() {
+    fn assemble_builtin_input_with_dataset_only_config() {
         let bench = qt::config::BuiltinBenchmarkConfig {
             type_: "builtin".to_owned(),
             samples: None,
+            dataset: "hf://quantiles/PubMedQA".to_owned(),
             model: None,
             max_workers: None,
         };
         let (input, _) = super::assemble_builtin_input(Some(&bench), None);
-        assert!(input.is_none());
+        let parsed: serde_json::Value = serde_json::from_str(&input.unwrap()).unwrap();
+        assert_eq!(parsed["dataset"], "hf://quantiles/PubMedQA");
     }
 
     /// When there is no config section at all and no CLI `--input`, builtin runs should
diff --git a/cli/src/config.rs b/cli/src/config.rs
index 0da4a18..2afa4b8 100644
--- a/cli/src/config.rs
+++ b/cli/src/config.rs
@@ -76,6 +76,8 @@ pub struct BuiltinBenchmarkConfig {
     pub type_: String,
     /// Number of samples (rows) to evaluate.
     pub samples: Option<usize>,
+    /// Dataset source for this benchmark.
+    pub dataset: String,
     /// Which model sampler to use for this benchmark.
     pub model: Option<Sampler>,
     /// Maximum concurrent workers for this benchmark.
@@ -159,6 +161,7 @@ mod tests {
     fn deserialize_builtin_without_type() {
         let toml = r#"
             [benchmarks.demo]
+            dataset = "hf://quantiles/demo"
             samples = 10
         "#;
         let config: WorkspaceConfig = toml::from_str(toml).unwrap();
@@ -167,6 +170,7 @@ mod tests {
         if let BenchmarkConfig::Builtin(b) = bench {
             assert_eq!(b.type_, "builtin");
             assert_eq!(b.samples, Some(10));
+            assert_eq!(b.dataset, "hf://quantiles/demo");
             assert!(b.model.is_none());
         }
     }
@@ -177,11 +181,25 @@ mod tests {
             [benchmarks.demo]
             type = "builtin"
             samples = 5
+            dataset = "hf://quantiles/demo"
             model = "openai:gpt-4"
         "#;
         let config: WorkspaceConfig = toml::from_str(toml).unwrap();
         let bench = config.benchmarks.get("demo").unwrap();
         assert!(matches!(bench, BenchmarkConfig::Builtin(_)));
+        if let BenchmarkConfig::Builtin(b) = bench {
+            assert_eq!(b.dataset, "hf://quantiles/demo");
+        }
+    }
+
+    #[test]
+    fn builtin_requires_dataset_field() {
+        let toml = r#"
+            [benchmarks.demo]
+            samples = 5
+        "#;
+        let result: Result<WorkspaceConfig, _> = toml::from_str(toml);
+        assert!(result.is_err(), "builtin should require dataset field");
     }
 
     #[test]
diff --git a/cli/src/dataset/mod.rs b/cli/src/dataset/mod.rs
index b740fd8..5242ff2 100644
--- a/cli/src/dataset/mod.rs
+++ b/cli/src/dataset/mod.rs
@@ -3,7 +3,7 @@ pub mod hf_client;
 
 use crate::dataset::cache::DatasetCache;
 use crate::dataset::hf_client::HuggingFaceClient;
-use anyhow::{Context, Result};
+use anyhow::{Context, Result, bail};
 use serde::Serialize;
 use serde_json::Value;
 
@@ -145,3 +145,60 @@ impl DatasetManager {
         splits.first().cloned().context("dataset has no splits")
     }
 }
+
+/// Resolve a configured Hugging Face dataset source to the dataset ID expected
+/// by the existing Hugging Face download client.
+pub fn resolve_hf_dataset_source(source: &str) -> Result<&str> {
+    if let Some(dataset_id) = source
+        .strip_prefix("hf://")
+        .or_else(|| source.strip_prefix("huggingface://"))
+    {
+        if dataset_id.is_empty() {
+            bail!("dataset source `{source}` is missing a Hugging Face dataset id");
+        }
+        Ok(dataset_id)
+    } else if source.contains("://") {
+        bail!("unsupported dataset source `{source}`; expected `hf://...` or `huggingface://...`");
+    } else {
+        bail!("dataset source `{source}` is missing required `hf://` or `huggingface://` prefix");
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::resolve_hf_dataset_source;
+
+    #[test]
+    fn resolve_hf_dataset_source_strips_hf_prefix() {
+        assert_eq!(
+            resolve_hf_dataset_source("hf://quantiles/PubMedQA").unwrap(),
+            "quantiles/PubMedQA"
+        );
+    }
+
+    #[test]
+    fn resolve_hf_dataset_source_strips_huggingface_prefix() {
+        assert_eq!(
+            resolve_hf_dataset_source("huggingface://quantiles/PubMedQA").unwrap(),
+            "quantiles/PubMedQA"
+        );
+    }
+
+    #[test]
+    fn resolve_hf_dataset_source_rejects_other_prefixes() {
+        let err = resolve_hf_dataset_source("s3://bucket/dataset").unwrap_err();
+        assert!(
+            err.to_string()
+                .contains("unsupported dataset source `s3://bucket/dataset`")
+        );
+    }
+
+    #[test]
+    fn resolve_hf_dataset_source_requires_prefix() {
+        let err = resolve_hf_dataset_source("quantiles/PubMedQA").unwrap_err();
+        assert!(
+            err.to_string()
+                .contains("missing required `hf://` or `huggingface://` prefix")
+        );
+    }
+}
diff --git a/cli/src/server/mod.rs b/cli/src/server/mod.rs
index 40b9631..94d9fbd 100644
--- a/cli/src/server/mod.rs
+++ b/cli/src/server/mod.rs
@@ -15,7 +15,7 @@ use sea_orm::DatabaseConnection;
 use serde_json::json;
 use tokio::net::TcpListener;
 
-use crate::dataset::DatasetManager;
+use crate::dataset::{DatasetManager, resolve_hf_dataset_source};
 use crate::db::steps::{self, StepDecision};
 use crate::db::{self, DBUrl, SQLitePathURL};
 use crate::metrics_store::MetricsStore;
@@ -321,14 +321,7 @@ async fn dataset_batch(
 }
 
 fn parse_hf_source(source: &str) -> Result<&str, ApiError> {
-    source
-        .strip_prefix("huggingface://")
-        .or_else(|| source.strip_prefix("hf://"))
-        .ok_or_else(|| {
-            ApiError(anyhow::anyhow!(
-                "unsupported dataset source `{source}`; expected `huggingface://...` or `hf://...`"
-            ))
-        })
+    resolve_hf_dataset_source(source).map_err(ApiError)
 }
 
 struct ApiError(anyhow::Error);