diff --git a/studio/backend/plugins/data-designer-github-repo-seed/src/data_designer_github_repo_seed/scraper.py b/studio/backend/plugins/data-designer-github-repo-seed/src/data_designer_github_repo_seed/scraper.py
index 3c60cddaf9..d768fe37be 100644
--- a/studio/backend/plugins/data-designer-github-repo-seed/src/data_designer_github_repo_seed/scraper.py
+++ b/studio/backend/plugins/data-designer-github-repo-seed/src/data_designer_github_repo_seed/scraper.py
@@ -177,6 +177,7 @@ def scrape(cfg: ScrapeConfig, base_dir: Path):
             base_dir = base_dir,
             client = client,
             trial_limits = trial_limits,
+            light = True,
         )
         try:
             repo_meta = scraper.scrape_repo_meta()
diff --git a/studio/backend/plugins/data-designer-github-repo-seed/src/data_designer_github_repo_seed/scraper_impl/queries.py b/studio/backend/plugins/data-designer-github-repo-seed/src/data_designer_github_repo_seed/scraper_impl/queries.py
index b21744a885..9dc7613db5 100644
--- a/studio/backend/plugins/data-designer-github-repo-seed/src/data_designer_github_repo_seed/scraper_impl/queries.py
+++ b/studio/backend/plugins/data-designer-github-repo-seed/src/data_designer_github_repo_seed/scraper_impl/queries.py
@@ -273,6 +273,66 @@ query PRsPage($owner: String!, $name: String!, $first: Int!, $after: String) {
 """,
 )
 
+PRS_PAGE_QUERY_LIGHT = _q(
+    [F_ACTOR, F_LABEL],
+    """
+query PRsPageLight($owner: String!, $name: String!, $first: Int!, $after: String) {
+  repository(owner: $owner, name: $name) {
+    pullRequests(first: $first, after: $after, orderBy: {field: CREATED_AT, direction: ASC}) {
+      pageInfo { hasNextPage endCursor }
+      totalCount
+      nodes {
+        id databaseId number title body state isDraft
+        createdAt updatedAt closedAt mergedAt
+        url
+        author { ...ActorFields }
+        labels(first: 50) { nodes { ...LabelFields } }
+        comments(first: 30) {
+          totalCount
+          pageInfo { hasNextPage endCursor }
+          nodes {
+            id databaseId createdAt updatedAt url body
+            author { ...ActorFields }
+          }
+        }
+      }
+    }
+  }
+  rateLimit { cost remaining resetAt }
+}
+""",
+)
+
+ISSUES_PAGE_QUERY_LIGHT = _q(
+    [F_ACTOR, F_LABEL],
+    """
+query IssuesPageLight($owner: String!, $name: String!, $first: Int!, $after: String) {
+  repository(owner: $owner, name: $name) {
+    issues(first: $first, after: $after, orderBy: {field: CREATED_AT, direction: ASC}) {
+      pageInfo { hasNextPage endCursor }
+      totalCount
+      nodes {
+        id databaseId number title body state
+        createdAt updatedAt closedAt
+        url
+        author { ...ActorFields }
+        labels(first: 50) { nodes { ...LabelFields } }
+        comments(first: 30) {
+          totalCount
+          pageInfo { hasNextPage endCursor }
+          nodes {
+            id databaseId createdAt updatedAt url body
+            author { ...ActorFields }
+          }
+        }
+      }
+    }
+  }
+  rateLimit { cost remaining resetAt }
+}
+""",
+)
+
 ISSUE_COMMENTS_QUERY = _q(
     [F_ACTOR],
     """
diff --git a/studio/backend/plugins/data-designer-github-repo-seed/src/data_designer_github_repo_seed/scraper_impl/scraper.py b/studio/backend/plugins/data-designer-github-repo-seed/src/data_designer_github_repo_seed/scraper_impl/scraper.py
index 19973bdfe1..ab1255ca31 100644
--- a/studio/backend/plugins/data-designer-github-repo-seed/src/data_designer_github_repo_seed/scraper_impl/scraper.py
+++ b/studio/backend/plugins/data-designer-github-repo-seed/src/data_designer_github_repo_seed/scraper_impl/scraper.py
@@ -42,12 +42,17 @@ class RepoScraper:
         base_dir: Path,
         client: GitHubClient,
         trial_limits: Optional[Dict[str, int]] = None,
+        light: bool = False,
     ):
         self.owner = owner
         self.name = name
         self.base_dir = base_dir
         self.client = client
         self.trial_limits = trial_limits or {}
+        # When light=True, use trimmed GraphQL queries (no reviewThreads,
+        # reviews, commits, timelineItems, files) so PR pages can be much
+        # larger without blowing GitHub's node-count ceiling.
+        self.light = light
         self.repo_dir = base_dir / f"{owner}__{name}"
         self.repo_dir.mkdir(parents = True, exist_ok = True)
         self.state = StateStore(base_dir / "state" / f"{owner}__{name}.json")
@@ -116,7 +121,8 @@ class RepoScraper:
             return 0
         total_new = 0
         page = 0
-        per_page = 15  # conservative for heavy nested query
+        # Light query skips heavy nested fields; safe at 50 per page.
+        per_page = 50 if self.light else 15
         while True:
             page += 1
             vars_ = {
@@ -125,7 +131,8 @@ class RepoScraper:
                 "first": per_page,
                 "after": cursor,
             }
-            data = self.client.graphql(Q.ISSUES_PAGE_QUERY, vars_)
+            query = Q.ISSUES_PAGE_QUERY_LIGHT if self.light else Q.ISSUES_PAGE_QUERY
+            data = self.client.graphql(query, vars_)
             self._log_rate("issues", data)
             repo = (data.get("data") or {}).get("repository") or {}
             issues = repo.get("issues") or {}
@@ -134,15 +141,20 @@ class RepoScraper:
                 it["_owner"] = self.owner
                 it["_repo"] = self.name
                 it["_fetchedAt"] = ts()
-                # Paginate nested comments/timeline if more exist
-                if it.get("comments", {}).get("pageInfo", {}).get("hasNextPage"):
-                    self._paginate_issue_comments(
-                        it["number"], it["comments"]["pageInfo"]["endCursor"]
-                    )
-                if it.get("timelineItems", {}).get("pageInfo", {}).get("hasNextPage"):
-                    self._paginate_issue_timeline(
-                        it["number"], it["timelineItems"]["pageInfo"]["endCursor"]
-                    )
+                if not self.light:
+                    if it.get("comments", {}).get("pageInfo", {}).get("hasNextPage"):
+                        self._paginate_issue_comments(
+                            it["number"], it["comments"]["pageInfo"]["endCursor"]
+                        )
+                    if (
+                        it.get("timelineItems", {})
+                        .get("pageInfo", {})
+                        .get("hasNextPage")
+                    ):
+                        self._paginate_issue_timeline(
+                            it["number"],
+                            it["timelineItems"]["pageInfo"]["endCursor"],
+                        )
                 if self.writers[key].write(it):
                     total_new += 1
             info = issues.get("pageInfo") or {}
@@ -217,7 +229,10 @@ class RepoScraper:
             return 0
         total_new = 0
         page = 0
-        per_page = 3  # PR query is heavy; keep small so huge PRs don't OOM GraphQL
+        # Heavy nested PR query is capped at 3 per page (GitHub node-count
+        # ceiling); light query skips reviewThreads/reviews/commits/etc and
+        # can safely go to 25 per page.
+        per_page = 25 if self.light else 3
         while True:
             page += 1
             vars_ = {
@@ -226,7 +241,8 @@ class RepoScraper:
                 "first": per_page,
                 "after": cursor,
             }
-            data = self.client.graphql(Q.PRS_PAGE_QUERY, vars_)
+            query = Q.PRS_PAGE_QUERY_LIGHT if self.light else Q.PRS_PAGE_QUERY
+            data = self.client.graphql(query, vars_)
             self._log_rate("prs", data)
             repo = (data.get("data") or {}).get("repository") or {}
             prs = repo.get("pullRequests") or {}
@@ -236,24 +252,35 @@ class RepoScraper:
                 pr["_repo"] = self.name
                 pr["_fetchedAt"] = ts()
                 num = pr["number"]
-                if pr.get("comments", {}).get("pageInfo", {}).get("hasNextPage"):
-                    self._paginate_pr_comments(
-                        num, pr["comments"]["pageInfo"]["endCursor"]
-                    )
-                if pr.get("timelineItems", {}).get("pageInfo", {}).get("hasNextPage"):
-                    self._paginate_pr_timeline(
-                        num, pr["timelineItems"]["pageInfo"]["endCursor"]
-                    )
-                if pr.get("commits", {}).get("pageInfo", {}).get("hasNextPage"):
-                    self._paginate_pr_commits(
-                        num, pr["commits"]["pageInfo"]["endCursor"]
-                    )
-                if pr.get("files", {}).get("pageInfo", {}).get("hasNextPage"):
-                    self._paginate_pr_files(num, pr["files"]["pageInfo"]["endCursor"])
-                if pr.get("reviewThreads", {}).get("pageInfo", {}).get("hasNextPage"):
-                    self._paginate_pr_review_threads(
-                        num, pr["reviewThreads"]["pageInfo"]["endCursor"]
-                    )
+                if not self.light:
+                    if pr.get("comments", {}).get("pageInfo", {}).get("hasNextPage"):
+                        self._paginate_pr_comments(
+                            num, pr["comments"]["pageInfo"]["endCursor"]
+                        )
+                    if (
+                        pr.get("timelineItems", {})
+                        .get("pageInfo", {})
+                        .get("hasNextPage")
+                    ):
+                        self._paginate_pr_timeline(
+                            num, pr["timelineItems"]["pageInfo"]["endCursor"]
+                        )
+                    if pr.get("commits", {}).get("pageInfo", {}).get("hasNextPage"):
+                        self._paginate_pr_commits(
+                            num, pr["commits"]["pageInfo"]["endCursor"]
+                        )
+                    if pr.get("files", {}).get("pageInfo", {}).get("hasNextPage"):
+                        self._paginate_pr_files(
+                            num, pr["files"]["pageInfo"]["endCursor"]
+                        )
+                    if (
+                        pr.get("reviewThreads", {})
+                        .get("pageInfo", {})
+                        .get("hasNextPage")
+                    ):
+                        self._paginate_pr_review_threads(
+                            num, pr["reviewThreads"]["pageInfo"]["endCursor"]
+                        )
                 if self.writers[key].write(pr):
                     total_new += 1
             info = prs.get("pageInfo") or {}
diff --git a/studio/frontend/src/features/data-recipes/learning-recipes/github-support-bot.json b/studio/frontend/src/features/data-recipes/learning-recipes/github-support-bot.json
index 2f3c911984..385554ab31 100644
--- a/studio/frontend/src/features/data-recipes/learning-recipes/github-support-bot.json
+++ b/studio/frontend/src/features/data-recipes/learning-recipes/github-support-bot.json
@@ -19,7 +19,8 @@
         "provider": "Local Model",
         "inference_parameters": {
           "temperature": 0.4,
-          "max_tokens": 1500
+          "max_tokens": 800,
+          "max_parallel_requests": 1
         }
       }
     ],
@@ -52,7 +53,8 @@
         "name": "support_answer",
         "drop": false,
         "model_alias": "model_1",
-        "prompt": "You are the Unsloth support assistant. Answer the user's question grounded in how Unsloth actually works. Produce structured JSON.\n\nSource issue/PR:\n- Repo: {{ repo }}\n- Title: {{ title }}\n- URL: {{ url }}\n- State: {{ state }}\n- Labels: {{ labels }}\n\nNormalized user question:\n{{ normalized_question }}\n\nRules:\n- `answer` is 3-8 sentences of Markdown. Include a python code block when showing usage.\n- `diagnosis_questions` is 1-4 extra questions the user should answer if the info is insufficient (versions, GPU, full traceback). Empty list if the answer is complete.\n- `cites` is a list of references the answer is grounded in (doc paths, GitHub URLs). Always include the source `url` above.\n- `confidence` is one of `high` / `medium` / `low`. Use `low` when ambiguous or out of scope.\n- Never invent APIs. If uncertain, set confidence=`low` and ask diagnosis_questions rather than fabricating.",
+        "prompt": "You are writing one training row for an Unsloth support bot. Produce structured JSON grounded in the GitHub thread.\n\nSource issue / PR:\n- Repo: {{ repo }}\n- Title: {{ title }}\n- URL: {{ url }}\n- State: {{ state }}\n- Labels: {{ labels }}\n- Body: {{ body }}\n- First comments: {{ comments }}\n\nNormalized user question:\n{{ normalized_question }}\n\nRules:\n- `answer`: 80-200 words of Markdown grounded in the thread. Cite the source URL at least once inline as `[source: {{ url }}]`.\n- When the thread is procedural (install, upgrade, fix), include a short ```bash or ```python code block if one appears in the thread.\n- Name at least one concrete symbol (function, class, flag, env var, or file path) from the thread when available.\n- Never recommend `rm -rf`, force push, or other destructive commands without a warning.\n- No em-dashes, no emojis, no AI-disclaimer phrases. Only cite URLs / paths that appear in the thread.\n- `diagnosis_questions`: 1-4 follow-ups when the thread is missing info (versions, GPU, traceback). Empty list if the answer is complete.\n- `cites`: URLs / file paths actually used. Always include `{{ url }}`.\n- `confidence`: `high` / `medium` / `low`. Use `low` when ambiguous or out of scope.",
+        "system_prompt": "You write grounded Unsloth support answers. Faithful to the thread, no invented facts, no em-dashes, no emojis, no AI-disclaimer phrases.",
         "output_format": {
           "type": "object",
           "properties": {
@@ -87,7 +89,7 @@
         "width": 400,
         "node_type": "markdown_note",
         "name": "note_1",
-        "markdown": "### GitHub Support Bot\nReal GitHub data to synthetic Q&A pairs for support-bot fine-tuning.\n\n**1. Click the Source Data node (top-left of the canvas) to enter:**\n- **Repos**: one `owner/name` per line\n- **GitHub token**: your `GH_TOKEN` (or leave blank to use the server's env var)\n- Item types (issues / pulls / commits) and per-resource limit\n\n**2. Click `Run` below and set the number of rows** (defaults to 10 for a quick test).\n\nThe built-in `github_repo` seed reader does rate-limit-aware GraphQL scraping. Each scraped item becomes a row with `title`, `body`, `comments`, `labels`, etc. Two LLM blocks turn it into `{normalized_question, support_answer}`. Answers are structured JSON with `answer` / `diagnosis_questions` / `cites` / `confidence`.",
+        "markdown": "### GitHub Support Bot\nReal GitHub data to synthetic Q&A pairs for support-bot fine-tuning.\n\n**Click `Run` below to generate 10 sample rows.** Defaults already point at `unslothai/unsloth` + `unslothai/unsloth-zoo`, use the server's `GH_TOKEN` env var, and run the bundled local model.\n\n**Upgrade to production**\n- Swap `unsloth/gemma-4-E2B-it-GGUF` for a larger model in the model_config node.\n- Replace the demo prompts on `normalized_question` / `support_answer` with the canonical codex pattern below (see Note 3).\n- Raise the seed `limit` from 100 to `0` (All) for a full backfill.\n- Raise `max_parallel_requests` back to 4 once your inference server can handle it.",
         "note_color": "#E0F2FE",
         "note_opacity": "35"
       },
@@ -109,7 +111,7 @@
         "width": 400,
         "node_type": "markdown_note",
         "name": "note_3",
-        "markdown": "The **answer** block enforces a JSON shape that maps directly onto the grounded RAG answerer format.\n\nFields:\n- `answer` (Markdown with code)\n- `diagnosis_questions` (follow-ups)\n- `cites` (doc + GitHub URLs)\n- `confidence` (high/medium/low)\n\nThis is key: uniform schema = less cleanup later.",
+        "markdown": "The **answer** block produces `{answer, diagnosis_questions, cites, confidence}`, ready to feed into the grounded RAG answerer.\n\n**Demo default**: 80-200 word answer, one inline `[source: <url>]` cite, `max_parallel_requests=1` so a small local model stays stable.\n\n**Production prompt (paste in):**\n- Require 150-300 word answers with 2+ citations.\n- Enforce named symbols (function / flag / env var / file path).\n- Code fences for every procedural step.\n- Use the resolved-issue / merged-PR template: `What changed / Fix location / Minimum repro`.\n- Reject rows containing em-dashes, emojis, or AI-disclaimer phrases.\n\nSee the Unsloth SupportBot dataset card for the full codex prompt we used to train `Gemma-4-Unsloth-Bot`.",
         "note_color": "#E0F2FE",
         "note_opacity": "35"
       }