Update logbook: Text-to-SQL Post-Training

logbook.js

@@ -455,7 +455,7 @@
           `<a class="embed-title" href="${esc(u)}" target="_blank" rel="noopener">${esc(id)}</a>` +
           `<a class="embed-open" href="${esc(u)}" target="_blank" rel="noopener">Open ↗</a>` +
           `</div>` +
-          `<iframe class="embed-frame" src="https://${sub}.hf.space" loading="lazy" ` +
+          `<iframe class="embed-frame" src="https://${sub}.hf.space/?sidebar=hidden&navbar=hidden" loading="lazy" ` +
           `allow="clipboard-read; clipboard-write; fullscreen"></iframe>`;
       },
     },

logbook.json

@@ -3,18 +3,54 @@
   "title": "Text-to-SQL Post-Training",
   "emoji": "🎯",
   "space_id": "abidlabs/text2sql-logbook",
-  "updated_at": "2026-07-02T06:19:56+00:00",
+  "updated_at": "2026-07-02T06:24:36+00:00",
   "root": {
     "slug": "index",
     "title": "Text-to-SQL Post-Training",
     "file": "pages/index.md",
     "children": [
+      {
+        "slug": "build-execution-accuracy-eval-harness",
+        "title": "Build execution-accuracy eval harness",
+        "file": "pages/build-execution-accuracy-eval-harness/page.md",
+        "children": []
+      },
+      {
+        "slug": "zero-shot-baselines-across-open-models",
+        "title": "Zero-shot baselines across open models",
+        "file": "pages/zero-shot-baselines-across-open-models/page.md",
+        "children": []
+      },
+      {
+        "slug": "clean-data-dedup-dialect-filtering",
+        "title": "Clean data: dedup + dialect filtering",
+        "file": "pages/clean-data-dedup-dialect-filtering/page.md",
+        "children": []
+      },
+      {
+        "slug": "qlora-sft-baseline",
+        "title": "QLoRA SFT baseline",
+        "file": "pages/qlora-sft-baseline/page.md",
+        "children": []
+      },
+      {
+        "slug": "lr-lora-rank-sweep",
+        "title": "LR & LoRA-rank sweep",
+        "file": "pages/lr-lora-rank-sweep/page.md",
+        "children": []
+      },
       {
         "slug": "prompt-format-ablation-chat-vs-completion",
         "title": "Prompt format ablation (chat vs completion)",
         "file": "pages/prompt-format-ablation-chat-vs-completion/page.md",
         "children": []
       },
+      {
+        "slug": "synthetic-data-augmentation-self-instruct",
+        "title": "Synthetic data augmentation (self-instruct)",
+        "file": "pages/synthetic-data-augmentation-self-instruct/page.md",
+        "children": []
+      },
       {
         "slug": "add-spider-wikisql-to-the-eval-suite",
         "title": "Add Spider + WikiSQL to the eval suite",
@@ -27,6 +63,12 @@
         "file": "pages/curriculum-order-by-join-complexity/page.md",
         "children": []
       },
+      {
+        "slug": "distill-from-a-larger-open-model",
+        "title": "Distill from a larger open model",
+        "file": "pages/distill-from-a-larger-open-model/page.md",
+        "children": []
+      },
       {
         "slug": "long-context-schema-eval-32k",
         "title": "Long-context schema eval @32k",
@@ -56,48 +98,6 @@
         "title": "Final model card + release",
         "file": "pages/final-model-card-release/page.md",
         "children": []
-      },
-      {
-        "slug": "build-execution-accuracy-eval-harness",
-        "title": "Build execution-accuracy eval harness",
-        "file": "pages/build-execution-accuracy-eval-harness/page.md",
-        "children": []
-      },
-      {
-        "slug": "zero-shot-baselines-across-open-models",
-        "title": "Zero-shot baselines across open models",
-        "file": "pages/zero-shot-baselines-across-open-models/page.md",
-        "children": []
-      },
-      {
-        "slug": "clean-data-dedup-dialect-filtering",
-        "title": "Clean data: dedup + dialect filtering",
-        "file": "pages/clean-data-dedup-dialect-filtering/page.md",
-        "children": []
-      },
-      {
-        "slug": "qlora-sft-baseline",
-        "title": "QLoRA SFT baseline",
-        "file": "pages/qlora-sft-baseline/page.md",
-        "children": []
-      },
-      {
-        "slug": "lr-lora-rank-sweep",
-        "title": "LR & LoRA-rank sweep",
-        "file": "pages/lr-lora-rank-sweep/page.md",
-        "children": []
-      },
-      {
-        "slug": "synthetic-data-augmentation-self-instruct",
-        "title": "Synthetic data augmentation (self-instruct)",
-        "file": "pages/synthetic-data-augmentation-self-instruct/page.md",
-        "children": []
-      },
-      {
-        "slug": "distill-from-a-larger-open-model",
-        "title": "Distill from a larger open model",
-        "file": "pages/distill-from-a-larger-open-model/page.md",
-        "children": []
       }
     ]
   }

logbook.md

@@ -27,29 +27,13 @@
 | planned | [CPU latency & throughput](#/cpu-latency-throughput) | to assign |
 | planned | [Final model card + release](#/final-model-card-release) | Ana |
 
-# Prompt format ablation (chat vs completion)
-
-# Add Spider + WikiSQL to the eval suite
-
-# Curriculum: order by join complexity
-
-# Long-context schema eval @32k
-
-# Full fine-tune vs LoRA comparison
-
-# Error taxonomy & failure analysis
-
-# CPU latency & throughput
-
-# Final model card + release
-
 # Build execution-accuracy eval harness
 
 ---
 
 ### Harness: execution accuracy over SQLite
 
-`Jul 02, 2026 · 06:19 UTC`
+`Jul 02, 2026 · 06:24 UTC`
 
 Execution accuracy is the right metric: exact string match is near-zero because the model writes semantically-equivalent but syntactically-varied SQL. The harness builds an in-memory SQLite DB from each example's schema, runs gold and predicted queries, and compares result sets (order-aware only when the gold has ORDER BY).
 
@@ -82,7 +66,7 @@ def execution_accuracy(preds, golds, schemas):
 
 ### Baselines: 28.9% best zero-shot
 
-`Jul 02, 2026 · 06:19 UTC`
+`Jul 02, 2026 · 06:24 UTC`
 
 Zero-shot execution accuracy on the 800-example held-out set. Instruct variants lead; the 1.5B instruct model is the best base to fine-tune from.
 
@@ -104,7 +88,7 @@ Target to beat with SFT: **28.9%**.
 
 ### Data: 42k clean SQLite-executable examples
 
-`Jul 02, 2026 · 06:19 UTC`
+`Jul 02, 2026 · 06:24 UTC`
 
 Filtered the training set to examples whose gold query executes cleanly in SQLite (~78% do; the rest use non-SQLite dialects), then deduped against the eval prompts. Final training set: 42k examples.
 
@@ -114,7 +98,7 @@ Filtered the training set to examples whose gold query executes cleanly in SQLit
 
 ### QLoRA baseline: 51.3% exec acc
 
-`Jul 02, 2026 · 06:19 UTC`
+`Jul 02, 2026 · 06:24 UTC`
 
 First SFT pass: QLoRA (r=16) on Qwen2.5-1.5B-Instruct, 3 epochs, completion-only loss. Execution accuracy 28.9% → **51.3%**. Live metrics on the Trackio dashboard.
 
@@ -146,20 +130,22 @@ if __name__ == "__main__":
 
 ### Sweep: r=16, lr=5e-4 wins
 
-`Jul 02, 2026 · 06:19 UTC`
+`Jul 02, 2026 · 06:24 UTC`
 
 Swept learning rate {1e-4, 2e-4, 5e-4} × rank {8, 16, 32}. r=16 / lr=5e-4 is the clear winner; r=8 underfits and lr>5e-4 destabilizes late in training.
 
 - media/lr_rank_sweep.png
 - https://huggingface.co/spaces/abidlabs/gemma-text2sql-trackio
 
+# Prompt format ablation (chat vs completion)
+
 # Synthetic data augmentation (self-instruct)
 
 ---
 
 ### Synth data: +3.1% exec acc (early)
 
-`Jul 02, 2026 · 06:19 UTC`
+`Jul 02, 2026 · 06:24 UTC`
 
 Generating extra (question, SQL) pairs by prompting a larger open model on real schemas, keeping only pairs whose SQL executes. Running as an HF Job; outputs land in a bucket. Early signal: +3.1% exec acc when mixed 1:4 with real data.
 
@@ -191,12 +177,26 @@ def augment(schemas, n_per_schema=8):
 - https://huggingface.co/jobs/abidlabs/6a45b02733c08a2c0dae0348
 - https://huggingface.co/buckets/abidlabs/jobs-artifacts
 
+# Add Spider + WikiSQL to the eval suite
+
+# Curriculum: order by join complexity
+
 # Distill from a larger open model
 
 ---
 
 ### Plan & hypothesis
 
-`Jul 02, 2026 · 06:19 UTC`
+`Jul 02, 2026 · 06:24 UTC`
 
 Plan: use the best open model as a teacher (rationale + SQL), distill into the 1.5B student. Hypothesis: closes most of the gap to the teacher at a fraction of the cost.
+
+# Long-context schema eval @32k
+
+# Full fine-tune vs LoRA comparison
+
+# Error taxonomy & failure analysis
+
+# CPU latency & throughput
+
+# Final model card + release

pages/build-execution-accuracy-eval-harness/page.md

@@ -3,8 +3,8 @@
 ---
 
 ### Harness: execution accuracy over SQLite
-<!-- entry ts=2026-07-02T06:19:56+00:00 -->
-`Jul 02, 2026 · 06:19 UTC`
+<!-- entry ts=2026-07-02T06:24:36+00:00 -->
+`Jul 02, 2026 · 06:24 UTC`
 
 Execution accuracy is the right metric: exact string match is near-zero because the model writes semantically-equivalent but syntactically-varied SQL. The harness builds an in-memory SQLite DB from each example's schema, runs gold and predicted queries, and compares result sets (order-aware only when the gold has ORDER BY).
 

pages/clean-data-dedup-dialect-filtering/page.md

@@ -3,7 +3,7 @@
 ---
 
 ### Data: 42k clean SQLite-executable examples
-<!-- entry ts=2026-07-02T06:19:56+00:00 -->
-`Jul 02, 2026 · 06:19 UTC`
+<!-- entry ts=2026-07-02T06:24:36+00:00 -->
+`Jul 02, 2026 · 06:24 UTC`
 
 Filtered the training set to examples whose gold query executes cleanly in SQLite (~78% do; the rest use non-SQLite dialects), then deduped against the eval prompts. Final training set: 42k examples.

pages/distill-from-a-larger-open-model/page.md

@@ -3,7 +3,7 @@
 ---
 
 ### Plan & hypothesis
-<!-- entry ts=2026-07-02T06:19:56+00:00 -->
-`Jul 02, 2026 · 06:19 UTC`
+<!-- entry ts=2026-07-02T06:24:36+00:00 -->
+`Jul 02, 2026 · 06:24 UTC`
 
 Plan: use the best open model as a teacher (rationale + SQL), distill into the 1.5B student. Hypothesis: closes most of the gap to the teacher at a fraction of the cost.

pages/lr-lora-rank-sweep/page.md

@@ -3,8 +3,8 @@
 ---
 
 ### Sweep: r=16, lr=5e-4 wins
-<!-- entry ts=2026-07-02T06:19:56+00:00 -->
-`Jul 02, 2026 · 06:19 UTC`
+<!-- entry ts=2026-07-02T06:24:36+00:00 -->
+`Jul 02, 2026 · 06:24 UTC`
 
 Swept learning rate {1e-4, 2e-4, 5e-4} × rank {8, 16, 32}. r=16 / lr=5e-4 is the clear winner; r=8 underfits and lr>5e-4 destabilizes late in training.
 

pages/qlora-sft-baseline/page.md

@@ -3,8 +3,8 @@
 ---
 
 ### QLoRA baseline: 51.3% exec acc
-<!-- entry ts=2026-07-02T06:19:56+00:00 -->
-`Jul 02, 2026 · 06:19 UTC`
+<!-- entry ts=2026-07-02T06:24:36+00:00 -->
+`Jul 02, 2026 · 06:24 UTC`
 
 First SFT pass: QLoRA (r=16) on Qwen2.5-1.5B-Instruct, 3 epochs, completion-only loss. Execution accuracy 28.9% → **51.3%**. Live metrics on the Trackio dashboard.
 

pages/synthetic-data-augmentation-self-instruct/page.md

@@ -3,8 +3,8 @@
 ---
 
 ### Synth data: +3.1% exec acc (early)
-<!-- entry ts=2026-07-02T06:19:56+00:00 -->
-`Jul 02, 2026 · 06:19 UTC`
+<!-- entry ts=2026-07-02T06:24:36+00:00 -->
+`Jul 02, 2026 · 06:24 UTC`
 
 Generating extra (question, SQL) pairs by prompting a larger open model on real schemas, keeping only pairs whose SQL executes. Running as an HF Job; outputs land in a bucket. Early signal: +3.1% exec acc when mixed 1:4 with real data.
 

pages/zero-shot-baselines-across-open-models/page.md

@@ -3,8 +3,8 @@
 ---
 
 ### Baselines: 28.9% best zero-shot
-<!-- entry ts=2026-07-02T06:19:56+00:00 -->
-`Jul 02, 2026 · 06:19 UTC`
+<!-- entry ts=2026-07-02T06:24:36+00:00 -->
+`Jul 02, 2026 · 06:24 UTC`
 
 Zero-shot execution accuracy on the 800-example held-out set. Instruct variants lead; the 1.5B instruct model is the best base to fine-tune from.