recipe(koelectra): add NSMC text classification recipe

[kujin66]

Add recipe support for daekeun-ml/koelectra-small-v3-nsmc, an ELECTRA sequence-classification checkpoint for Korean NSMC sentiment classification. This is a recipe-only L0 contribution with Goal L1 evidence on CPU for fp16 and w8a16. Baseline auto-config already builds the model, so the measurable delta is the checked-in catalog recipe plus verified w8a16 static QDQ quantization path.

1. Recipe path(s)

• examples/recipes/daekeun-ml_koelectra-small-v3-nsmc/text-classification_fp16_config.json
• examples/recipes/daekeun-ml_koelectra-small-v3-nsmc/text-classification_w8a16_config.json

2. README row

• Added | daekeun-ml/koelectra-small-v3-nsmc | text-classification | to examples/recipes/README.md.
• Updated total from 75 to 76.

3. Build output dir

• Baseline auto-config: temp/koelectra_baseline_build
• Recipe fp16: temp/koelectra_fp16_build
• Recipe w8a16 static QDQ: temp/koelectra_w8a16_quant_build2

4. Build log

• Baseline command: .\.venv\Scripts\python.exe -m winml.modelkit build -m daekeun-ml/koelectra-small-v3-nsmc -o temp/koelectra_baseline_build --ep cpu --device cpu --no-analyze --no-optimize --no-quant --no-compile --rebuild
  • Output contained Build complete in 47.0s, final artifact temp\koelectra_baseline_build\model.onnx.
  • Process exited by timeout after success due to repeated HuggingFace proxy retries against 127.0.0.1:9.
  • winml --version: winml, version 0.2.0.
  • Baseline/rebase base: origin/main = e0751f6dedfb45e2a1a177a7fff74a111b880514; review check git rev-list --count e0751f6dedfb45e2a1a177a7fff74a111b880514..origin/main = 0.
  • PR commit: 23b8522077fe9f4cc541b286421c30ad3c186349.
• fp16 recipe command: .\.venv\Scripts\python.exe -m winml.modelkit build -c examples/recipes/daekeun-ml_koelectra-small-v3-nsmc/text-classification_fp16_config.json -m daekeun-ml/koelectra-small-v3-nsmc -o temp/koelectra_fp16_build --ep cpu --device cpu --no-analyze --no-optimize --no-quant --no-compile --rebuild
  • Output contained Build complete in 55.5s, final artifact temp\koelectra_fp16_build\model.onnx.
• w8a16 recipe command: .\.venv\Scripts\python.exe -m winml.modelkit build -c examples/recipes/daekeun-ml_koelectra-small-v3-nsmc/text-classification_w8a16_config.json -m daekeun-ml/koelectra-small-v3-nsmc -o temp/koelectra_w8a16_quant_build2 --precision w8a16 --ep cpu --device cpu --no-analyze --no-optimize --no-compile --rebuild
  • Final artifact temp\koelectra_w8a16_quant_build2\model.onnx.
  • ONNX check: QuantizeLinear=426, DequantizeLinear=634, inputs input_ids/attention_mask/token_type_ids [1,512], output logits [1,2].

5. Appended findings

• Added Skill knowledge file model_knowledge/electra.json with finding electra-001.
• Finding captures fp16/w8a16 CPU evidence, export metadata, QDQ node counts, and gotchas.

6. Optimum-coverage probe

{
  "vendor": [],
  "after_winml": ["feature-extraction", "fill-mask", "multiple-choice", "question-answering", "text-classification", "token-classification"],
  "added_by_winml": ["feature-extraction", "fill-mask", "multiple-choice", "question-answering", "text-classification", "token-classification"],
  "verdict": "WINML-ONLY"
}

Model config probe:

{
  "model_type": "electra",
  "architectures": ["ElectraForSequenceClassification"],
  "vocab_size": 35000,
  "max_position_embeddings": 512,
  "type_vocab_size": 2
}

7. Claimed (Effort, Goal, Outcome)

{
  "model_id": "daekeun-ml/koelectra-small-v3-nsmc",
  "model_type": "electra",
  "effort": "L0",
  "goal_ceiling": "L1",
  "outcome": "L0",
  "target_eps": ["cpu"],
  "catalog_gate": {
    "baseline_build": "PASS",
    "verdict": "catalog-needs-numbers",
    "file_pr": true
  }
}

8. Goal-ladder verdict table

{
  "deliverable_ref": "daekeun-ml/koelectra-small-v3-nsmc",
  "ladder": [
    {
      "tier": "L0",
      "verdict": "PASS",
      "evidence": "Baseline auto-config and recipe builds produced loadable ONNX artifacts. fp16 model.onnx: IR=8, opset=17, inputs input_ids/attention_mask/token_type_ids [1,512], output logits [1,2]. w8a16 model.onnx: inputs input_ids/attention_mask/token_type_ids [1,512], output logits [1,2], QuantizeLinear=426, DequantizeLinear=634."
    },
    {
      "tier": "L1",
      "verdict": "PASS",
      "evidence": "CPU providers snapshot: ['DmlExecutionProvider', 'CPUExecutionProvider']. fp16 CPU perf mean 53.371 ms, throughput 18.74 samples/sec, RSS total delta 95.37 MB. w8a16 CPU perf mean 128.133 ms, throughput 7.8 samples/sec, RSS total delta 88.58 MB."
    }
  ],
  "ceiling_reached": "L1",
  "ceiling_downgraded": false,
  "new_verdict_shapes": []
}

9. Methodology-evolution declaration

• Methodology friction observed: validating a quantized recipe with a plain --device cpu --ep cpu build can resolve saved quant to null, so quantized recipes need artifact-level QDQ validation.
• Captured in local Skill model knowledge as model_knowledge/electra.json finding electra-001.
• Captured in local Skill methodology knowledge as skill_meta/findings.json finding _meta-057, with paired edit in agents/reviewer.md requiring QDQ node / saved quant-config validation for _w8a16_ and _w8a8_ recipes.
• The Skill Lane A files are not included in this Lane B winml-cli model PR, preserving the model PR's recipe-only diff scope.

Additional notes

• PR created and marked ready for review: recipe(koelectra): add NSMC text classification recipe #1144
• Branch was rebased onto origin/main e0751f6dedfb45e2a1a177a7fff74a111b880514 before commit 23b8522077fe9f4cc541b286421c30ad3c186349.

[kujin66]

Reviewer verdict: APPROVE fallback comment / Lane B model PR is ready for maintainer review.

Checked items:

• PR body contains the 9-item adding-model-support report and no longer has stale gh-auth or rebase blocker text.
• Diff scope is L0 recipe-only: README plus two KOElectra recipe JSON files.
• Branch is based on current origin/main e0751f6dedfb45e2a1a177a7fff74a111b880514; git rev-list --count e0751f6dedfb45e2a1a177a7fff74a111b880514..origin/main returned 0.
• CI is green: Analyze, CodeQL, CLA, lint, and all test jobs passed.
• Goal L1 evidence is present for CPU: fp16 perf and w8a16 perf are reported, and the w8a16 artifact has nonzero QDQ nodes (QuantizeLinear=426, DequantizeLinear=634).
• PR is mergeable according to GitHub. GitHub-side approval metadata could not be set because the same account cannot approve its own pull request.

Open process note:

• Methodology friction was observed around quantized-recipe validation requiring artifact-level QDQ checks. It is captured in local Skill/model knowledge as model_knowledge/electra.json / electra-001 and in local Skill methodology knowledge as skill_meta/findings.json / _meta-057, with paired agents/reviewer.md checklist text.
• Skill Lane A files are intentionally not included in this Lane B winml-cli PR so the model diff remains recipe-only.

I attempted gh pr review --approve; GitHub rejected it with Review Can not approve your own pull request, so this structured APPROVE comment is the _meta-035 fallback verdict.