docs: clarify SWE-bench dataset downloads

fern/versions/latest/pages/data/prepare-validate.mdx

@@ -224,6 +224,37 @@ This validates your data and adds the `agent_ref` field to each row, routing sam
 
 ---
 
+## Download Missing Remote Datasets
+
+If a dataset in your config has a remote identifier, `ng_prepare_data` can download it before validation. Use `+should_download=true` and select the matching `+data_source`.
+
+For configs that use `gitlab_identifier`, such as the SWE agents `swebench_openhands.yaml` validation dataset, use the GitLab registry backend:
+
+```bash
+config_paths="responses_api_agents/swe_agents/configs/swebench_openhands.yaml,\
+responses_api_models/vllm_model/configs/vllm_model.yaml"
+
+ng_prepare_data "+config_paths=[$config_paths]" \
+    +output_dirpath=data/swebench_openhands \
+    +mode=train_preparation \
+    +should_download=true \
+    +data_source=gitlab
+```
+
+The same dataset can also be downloaded directly:
+
+```bash
+ng_download_dataset_from_gitlab \
+    +dataset_name=swebench_verified_for_sweagent_and_openhands \
+    +version=0.0.1 \
+    +artifact_fpath=swebench_verified_for_sweagent_and_openhands.jsonl \
+    +output_fpath=responses_api_agents/swe_agents/data/swebench_verified_for_sweagent_and_openhands.jsonl
+```
+
+Use `ng_download_dataset_from_hf` only for configs that use `huggingface_identifier`. A 401 from the Hugging Face downloader usually means the command is pointing at the wrong registry or you need access to a gated/private Hugging Face repo.
+
+---
+
 ## Validation Modes
 
 | Mode | Purpose | Validates |

fern/versions/latest/pages/reference/cli-commands.mdx

@@ -261,6 +261,16 @@ ng_prepare_data "+config_paths=[${config_paths}]" \
     +mode=example_validation
 ```
 
+To download missing datasets from a config before validation, add `+should_download=true` and set `+data_source` to the registry used by that dataset's identifier. For example, configs with `gitlab_identifier` use the GitLab dataset registry:
+
+```bash
+ng_prepare_data "+config_paths=[responses_api_agents/swe_agents/configs/swebench_openhands.yaml,responses_api_models/vllm_model/configs/vllm_model.yaml]" \
+    +output_dirpath=data/swebench_openhands \
+    +mode=train_preparation \
+    +should_download=true \
+    +data_source=gitlab
+```
+
 ---
 
 ## Dataset Registry - GitLab

responses_api_agents/swe_agents/README.md

@@ -221,6 +221,43 @@ apt install -y wget && cd /tmp && \
 apptainer --version
 ```
 
+### Download the SWE-bench Verified dataset
+
+The bundled `configs/swebench_openhands.yaml` validation dataset points at NeMo Gym's GitLab/MLflow dataset registry:
+
+```yaml
+jsonl_fpath: responses_api_agents/swe_agents/data/swebench_verified_for_sweagent_and_openhands.jsonl
+gitlab_identifier:
+  dataset_name: swebench_verified_for_sweagent_and_openhands
+  version: 0.0.1
+  artifact_fpath: swebench_verified_for_sweagent_and_openhands.jsonl
+```
+
+Download it with the GitLab dataset command, not the Hugging Face downloader:
+
+```bash
+ng_download_dataset_from_gitlab \
+    +dataset_name=swebench_verified_for_sweagent_and_openhands \
+    +version=0.0.1 \
+    +artifact_fpath=swebench_verified_for_sweagent_and_openhands.jsonl \
+    +output_fpath=responses_api_agents/swe_agents/data/swebench_verified_for_sweagent_and_openhands.jsonl
+```
+
+Or let `ng_prepare_data` fetch any missing datasets from the config:
+
+```bash
+config_paths="responses_api_agents/swe_agents/configs/swebench_openhands.yaml,\
+responses_api_models/vllm_model/configs/vllm_model.yaml"
+
+ng_prepare_data "+config_paths=[$config_paths]" \
+    +output_dirpath=data/swebench_openhands \
+    +mode=train_preparation \
+    +should_download=true \
+    +data_source=gitlab
+```
+
+The GitLab path requires the MLflow/GitLab registry credentials configured for your NeMo Gym environment. If those credentials are unavailable, ask your project admin for registry access or replace the config with a dataset you can access. `ng_download_dataset_from_hf` is only for configs that use `huggingface_identifier`, so it will not resolve this dataset name.
+
 ### Step 1 — configure the model
 
 In `env.yaml` at the NeMo-Gym root: