pull-hf-model: support datasets and spaces, not just models

Adds optional repo_type var (default "model"; valid: model/dataset/space)
that's forwarded to `hf download --repo-type` and threaded through the
verify step (cache-dir prefix tracks repo_type: models--, datasets--,
spaces--).

The playbook was previously model-only because hf download defaults
repo_type=model — pulling a dataset surfaced as a 404 against the wrong
HF API path. Now generic.

Smoke-tested 2026-05-13:
  scripts/elway ana-ml2 --playbook playbooks/pull-hf-model.yaml \
    --var hf_repo=Skywork/Skywork-Reward-Preference-80K-v0.2 \
    --var repo_type=dataset
  → 200 MB parquet cached at
    /tank/aimodels/huggingface/hub/datasets--Skywork--Skywork-Reward-Preference-80K-v0.2/

Filename is misleading now (playbook handles more than models); leaving
the rename for a future cleanup since pull-hf-model.yaml is referenced
from the llama-swap config.yaml's 2026-05-13 changelog entry.

playbooks/pull-hf-model.yaml

@@ -1,16 +1,17 @@
-# Pull a model from HuggingFace into ana-ml2's shared HF cache
-# (`/tank/aimodels/huggingface/`). Model-format-agnostic — handles GGUFs,
-# safetensors, or anything else HF hosts.
+# Pull a model or dataset from HuggingFace into ana-ml2's shared HF cache
+# (`/tank/aimodels/huggingface/`). Repo-type-agnostic — handles GGUFs,
+# safetensors, datasets, or anything else HF hosts.
 #
-# Consumers (llama-swap, vLLM, others) read from the same cache:
+# Consumers (llama-swap, vLLM, training jobs) read from the same cache:
 #   - llama-swap's config.yaml references entries via `-hf <user>/<repo>`
 #   - vLLM stacks reference via `--model <user>/<repo>` with HF_HOME mounted
+#   - datasets.load_dataset() reads from HF_HOME/datasets too
 #
 # This playbook does NOT edit any consumer config. Run it for the download,
 # then hand-edit the relevant `stacks/<consumer>/conf/*.yaml` and deploy
 # that consumer separately. Keeps the download mechanical and the
-# per-model params (ctx-size, quant choice, sampler defaults, etc.) where
-# they belong — in human-curated config.
+# per-asset params (ctx-size, quant choice, sampler defaults, split filters,
+# etc.) where they belong — in human-curated config.
 #
 # Usage:
 #
@@ -23,23 +24,37 @@
 #       --var hf_repo=mradermacher/Selene-1-Mini-Llama-3.1-8B-GGUF \
 #       --var allow_patterns='*Q6_K*'
 #
+#   # Dataset (preference data, eval set, etc.):
+#   scripts/elway ana-ml2 --playbook playbooks/pull-hf-model.yaml \
+#       --var hf_repo=Skywork/Skywork-Reward-Preference-80K-v0.2 \
+#       --var repo_type=dataset
+#
 # allow_patterns is forwarded to `hf download --include` — set it to filter
-# down to specific files (e.g. `*UD-Q6_K_XL*` or `*Q6_K*.gguf`). Omit to
+# down to specific files (e.g. `*UD-Q6_K_XL*` or `train.parquet`). Omit to
 # pull every file in the repo.
+#
+# repo_type — one of `model` (default), `dataset`, `space`. Affects both the
+# `hf download --repo-type` flag AND the cache dir prefix (`models--…`,
+# `datasets--…`, or `spaces--…`).
 
 vars:
   hf_repo: ""             # REQUIRED, e.g. unsloth/Qwen3.6-35B-A3B-GGUF
   allow_patterns: ""      # OPTIONAL, e.g. "*Q6_K*"
+  repo_type: "model"      # OPTIONAL, one of: model | dataset | space
   hf_cache_dir: /tank/aimodels/huggingface
 
 steps:
-  # ── Sanity: refuse to run with an empty hf_repo. ──────────────────────
-  - name: Check hf_repo is set
+  # ── Sanity: refuse to run with an empty hf_repo or an unknown repo_type. ──
+  - name: Check hf_repo + repo_type are valid
     shell: |
       [ -n "{{ hf_repo }}" ] || {
         echo "elway: --var hf_repo=<user>/<repo> is required" >&2
         exit 2
       }
+      case "{{ repo_type }}" in
+        model|dataset|space) ;;
+        *) echo "elway: repo_type must be one of: model, dataset, space (got '{{ repo_type }}')" >&2; exit 2 ;;
+      esac
     changed_when: "false"
 
   # ── Tooling: install `hf` CLI via pipx once, inject hf_transfer for
@@ -59,13 +74,13 @@ steps:
 
   # ── Pull. hf CLI is itself idempotent — re-runs only fetch missing
   #         blobs, so safe to invoke unconditionally. ─────────────────
-  - name: Pull {{ hf_repo }} into {{ hf_cache_dir }}
+  - name: Pull {{ hf_repo }} ({{ repo_type }}) into {{ hf_cache_dir }}
     shell: |
       export PATH="$HOME/.local/bin:$PATH"
       pat="{{ allow_patterns }}"
-      args=""
+      args="--repo-type {{ repo_type }}"
       if [ -n "$pat" ]; then
-        args="--include $pat"
+        args="$args --include $pat"
       fi
       HF_HOME={{ hf_cache_dir }} \
       HF_HUB_ENABLE_HF_TRANSFER=1 \
@@ -75,12 +90,14 @@ steps:
     # always-reporting-ok here; the verify step below confirms presence.
     changed_when: "false"
 
-  - name: Verify cache hit for {{ hf_repo }}
+  - name: Verify cache hit for {{ hf_repo }} ({{ repo_type }})
     shell: |
-      # hf normalizes user/repo to models--user--repo in the hub cache.
+      # hf normalizes user/repo to <kind>--user--repo in the hub cache.
+      # Cache-dir prefix tracks repo_type: models--, datasets--, spaces--.
+      prefix="{{ repo_type }}s--"
       slug=$(printf '%s' "{{ hf_repo }}" | tr '/' '~')
       slug=${slug//\~/--}
-      d="{{ hf_cache_dir }}/hub/models--${slug}"
+      d="{{ hf_cache_dir }}/hub/${prefix}${slug}"
       [ -d "$d" ] || { echo "elway: cache dir not found for {{ hf_repo }} at $d" >&2; exit 1; }
       n=$(find "$d/snapshots/" -mindepth 2 -maxdepth 2 \( -type f -o -type l \) 2>/dev/null | wc -l)
       [ "$n" -gt 0 ] || { echo "elway: snapshot dir exists but is empty under $d" >&2; exit 1; }
@@ -96,10 +113,11 @@ verify:
     shell: test -w {{ hf_cache_dir }}
     changed_when: "false"
 
-  - name: Model present in cache (post-pull)
+  - name: Repo present in cache (post-pull)
     shell: |
+      prefix="{{ repo_type }}s--"
       slug=$(printf '%s' "{{ hf_repo }}" | tr '/' '~')
       slug=${slug//\~/--}
-      d="{{ hf_cache_dir }}/hub/models--${slug}"
+      d="{{ hf_cache_dir }}/hub/${prefix}${slug}"
       test -d "$d" && [ -n "$(find $d/snapshots/ -mindepth 2 -maxdepth 2 \( -type f -o -type l \) 2>/dev/null | head -1)" ]
     changed_when: "false"