feat: initial release

README.md

@@ -0,0 +1,101 @@
+---
+language:
+- ko
+license: apache-2.0
+task_categories:
+- text-generation
+- question-answering
+tags:
+- korean
+- reasoning
+- chain-of-thought
+- verified
+- rejection-sampling
+- synthetic
+size_categories:
+- 1K<n<10K
+configs:
+- config_name: default
+  data_files:
+  - split: train
+    path: train.jsonl
+---
+
+# ko-verified-cot
+
+**2,016 short Korean reasoning traces whose final answer was verified against a ground-truth
+answer key.** Wrong reasoning was thrown away, not kept.
+
+[![GitHub](https://img.shields.io/badge/GitHub-kogemma--e2b-181717?logo=github)](https://github.com/waylake/kogemma-e2b)
+[![Model](https://img.shields.io/badge/model-KoGemma--E2B-4285F4)](https://huggingface.co/davek/KoGemma-E2B)
+
+## How it was built
+
+1. Sample multiple-choice questions from the **train split** of
+   [KMMLU](https://huggingface.co/datasets/HAERAE-HUB/KMMLU) (45 subjects) — test split is never
+   touched, so downstream evaluation stays clean.
+2. Ask the teacher (`ox-alpha-free`) for a **short** Korean chain of thought (3–4 sentences) ending
+   in `정답: X`. The teacher never sees the answer key.
+3. Extract the predicted letter and compare with the gold label. **Keep only matches.**
+4. Strip markdown, drop truncated traces, normalise the ending to `정답: X`.
+
+Teacher accuracy on those questions was **80–85%**; the same base model we later fine-tuned scored
+53% (train) / 32% (test). So the released traces are a strictly better-than-base subset.
+
+## Why short traces
+
+Small models (≤3B) get **worse** when trained on long/complex reasoning chains — see
+[*Small Models Struggle to Learn from Strong Reasoners*](https://arxiv.org/abs/2502.12143)
+(Qwen2.5-1.5B: 27.0 with long CoT vs 34.2 with short CoT). Traces here are deliberately capped at
+3–4 sentences to stay inside a small model's learnable range.
+
+## Schema
+
+```json
+{
+  "messages": [
+    {"role": "system", "content": "너는 한국어로 정확하게 추론하고 답하는 AI 비서다."},
+    {"role": "user", "content": "<질문>\n\nA. …\nB. …\nC. …\nD. …\n\n핵심 근거만 3~4문장으로 …"},
+    {"role": "assistant", "content": "<3~4문장 추론>\n\n정답: B"}
+  ],
+  "source": "ox-alpha-free"
+}
+```
+
+The user turn keeps a fixed instruction suffix so that training and evaluation prompts match
+exactly — misaligned prompt formats were the single largest source of measurement error in this
+project.
+
+## What we learned using it (honest results)
+
+Fine-tuning `gemma-4-E2B-it` (2.3B effective) on these traces mixed 1:1 with the model's own
+verified CoT (the *Mix Distillation* recipe):
+
+| Metric (KMMLU test, 600 questions) | base | after training |
+|---|---|---|
+| CoT accuracy | 0.3200 | 0.3250 |
+| CoT format failures | 48 (8.0%) | **27 (4.5%)** |
+
+Accuracy did **not** move outside sampling error; format compliance clearly improved. Verified
+teacher traces fix *how* a small model answers, not *how much it knows* — the bottleneck at 2.3B
+is capacity, not data.
+
+Also worth recording: measuring on 200 questions showed a fake +3.0 pp gain that vanished at 600.
+
+## Limitations
+
+- Multiple-choice reasoning only (KMMLU-style), Korean only.
+- Verified means "final answer matched"; individual reasoning steps are not step-verified.
+- Derived from KMMLU train questions — do not train and then claim KMMLU test gains without
+  checking split hygiene.
+
+## Citation
+
+```bibtex
+@misc{ko-verified-cot-2026,
+  title  = {ko-verified-cot: answer-verified short Korean reasoning traces},
+  author = {waylake},
+  year   = {2026},
+  url    = {https://huggingface.co/datasets/davek/ko-verified-cot}
+}
+```