Cheva123/jamjuri-mix7m-th

Jamjuri Mix7M

Clean Thai fine-tune mix for benchmark maxxing: 30,069 docs / 8,302,257 tokens (Qwen2.5 tokenizer, no special tokens).

Derived from SPAISS6F1/spai-ss6-llm-1b-thai-corpus — check upstream terms (medical/exam subsets carry their own licenses) before commercial use. Research/eval use.

Composition

Source	Docs	Tokens	Share
lst20 Thai QA	7,643	2,796,494	33.7%
exam QA	8,109	2,267,511	27.3%
synthetic_qa (10% sample)	1,324	1,200,296	14.5%
medical_qa, non-star (34% sample, CoT dropped, Q+A only)	5,579	800,210	9.6%
tourist instruction (9% sample)	2,454	600,278	7.2%
local_v2 instruction (9% sample)	3,641	500,035	6.0%
idioms (full)	1,152	113,001	1.4%
synonym (full)	167	24,432	0.3%

Cleaning

• Dropped <think>-containing rows (found zero — corpus already clean)
• Min 20 chars, exact-md5 dedupe 137,708 → 111,563 before sampling
• Stratified-random sample, seed 42

Format

mix7m.jsonl: one JSON per line — {"text": ..., "source": ..., "ntok": ...}. MANIFEST.json: exact counts.