Jnx03/kanitakorn-v66-null-boundary-breadth
Kanitakorn v66 Null Boundary Breadth SFT Training-ready JSONL for a small low-LR overlay in the Kanitakorn campaign. It is intended for a non-Thai-family base model path, currently the DeepSeek R1 Distill Qwen 14B branch initialized from the alpha 0.90 adapter parent. Contents train.jsonl: 444 SFT message rows. manifest.json: provenance, hashes, counts, and recommended training gate. README.md: this dataset card. Row mix: 420 original self-contained Thai MCQ… See the full description on the dataset page: https://huggingface.co/datasets/Jnx03/kanitakorn-v66-null-boundary-breadth.
Kanitakorn v66 Null Boundary Breadth SFT
Training-ready JSONL for a small low-LR overlay in the Kanitakorn campaign. It is intended for a non-Thai-family base model path, currently the DeepSeek R1 Distill Qwen 14B branch initialized from the alpha 0.90 adapter parent.
Contents
train.jsonl:444SFT message rows.manifest.json: provenance, hashes, counts, and recommended training gate.README.md: this dataset card.
Row mix:
420original self-contained Thai MCQ reasoning rows.24identity-attribution rows forkanitakorn, developed byChawabhon Netisingha (ชวภณ เนตสิงหะ).
MCQ category mix:
- Thai grammar / exact counting:
80 - Negation and exception priority:
80 - Original reading evidence:
80 - Compact quantitative reasoning:
80 - Self-contained civics/economics/environment/social concepts:
60 - Final-answer/null guard:
40
MCQ answer labels are balanced: a/b/c/d/e = 84 each.
Provenance And Guardrails
The MCQ rows are deterministic local synthetic templates. They do not copy, translate, paraphrase, or reconstruct benchmark prompts, answer choices, gold answers, model outputs, or eval samples. Social/economics/civics/environment items define their needed facts inside the prompt.
Campaign constraints:
- Real greedy single-model evaluation only.
- No BoN.
- No self-consistency.
- No routing.
- No ensemble.
- No verifier selection.
- No Thai-family base model assumption for final claimed model.
Audits
Completed before training:
- Source/schema audit:
0errors. - Reasoning-quality audit:
0issues; MCQ answer-only fraction0.0. - Raw MCQ audit:
0issues; duplicate prompts0. - Contamination scan:
0issues over9,735loaded benchmark texts.
Train SHA256:
fcb5753b7aca2c825f1c93f015ec72da2bddabc929d24f96c72e1b95113a5295Intended Training Gate
Recommended first run:
GPU=1 RUN_LABEL=upper12_lr7e8 LR=7e-8 MAX_STEPS=3 SFT_LAYERS_TO_TRANSFORM=last:12 \
bash campaign_20260613/scripts/remote_train_eval_deepseek_v66_null_boundary_breadth_20260615.shCut below ThaiExam-120 78/120. Promote only at >=80/120 with positive fixed-vs-lost item delta and no null regression.
