MC7ever/all-in-one-llm-corpus

🧬 All-in-One LLM Corpus

One repo. Four training stages. Everything a chat model needs — and nothing that sounds like a bot.

SFT → DPO → RL → persona polish → benchmark, with clean splits for each step, de-slopified in v2 so the model learns to talk, act, think and pass as a person.

🔢 Total rows	480,032 across 4 subsets × 4 splits
🎯 Subsets	sft · ct · rl · persona
✂️ Splits	train · validation · test · benchmark (all four in every subset)
📐 Split policy	~75% train / 25% eval ⇒ 1 eval row per 3 train rows
🚫 Leakage control	55,867 benchmark items never appear in train/validation/test
🧹 Dedup	525,109 raw blend rows → 321,785 unique (203,324 cross-version dupes removed), then v2 de-bot pass −19,874
🤖 De-bot v2	−10,665 bot/corporo targets · −2,334 purple-prose density · assistant-identity system lines stripped
🌍 Languages	language column on every row: en, zh, ja, ko, ru, de, uk, cs, sk (+ long tail)
🆕 v2 sources	OpenAssistant/oasst1 + CohereLabs/aya (human-written) · facebook/belebele (cs/sk/de/uk eval)
🤖 Models	model-agnostic chat format (LFM2.5, Qwen, Llama, Gemma, Mistral, …)

🆕 What's new in v2

• De-botting pass — every row's target text is screened against a combined bot/corporo phrase list (heretic noslop · nohumor markers + curated assistant-identity & corpo-speak patterns + DrRiceIO7/SlopReview-mined phrases). Assistant-identity system lines ("You are a helpful assistant.") are stripped, and prompts that interrogate AI identity are dropped. See Results → Filter quality for measured scores.
• Human-written language additions — German, Czech, Slovak and Ukrainian rows from OASST1 (crowd-written conversations) and Aya (human-annotated), deduped against the existing corpus and slop-filtered with the same rules.
• language column everywhere — langdetect-tagged on every row of every subset/split; filter to any language in one line.
• belebele benchmarks — reading-comprehension MCQ in deu_Latn, ces_Latn, slk_Latn, ukr_Cyrl added to sft.benchmark (eval-only; Czech & Slovak training data remains a documented gap — quality human-written cs/sk chat is not publicly available in bulk).
• Expanded card metadata — task categories, tags, full language table, results & scores.

Subsets

subset	what it's for	schema	train / validation / test / benchmark
sft	Supervised fine-tuning & continued SFT on blended chat data	messages, source, subset, language	271,491 / 20,145 / 20,161 / 55,867
ct	Continued training with DPO/ORPO preference pairs	prompt, system, chosen, rejected, source, language	46,813 / 5,240 / 5,196 / 5,229
rl	GRPO / verifiable-reward RL tasks	task_id, family, prompt, answer, verifier, tools, metadata, language	34,241 / 3,813 / 3,795 / 3,813
persona	Persona & style transfer SFT (system-prompt heavy)	system, user, assistant, source, language	3,171 / 352 / 352 / 353

Splits

The same four splits exist inside every subset:

• train — 75% of the subset. The only split you update weights on.
• validation — held-out, in-distribution: early stopping, mid-training loss / reward.
• test — untouched, in-distribution holdout: your final in-distribution number.
• benchmark — scored against ground truth, never trained on.
  • sft.benchmark = the 55,867-item benchmark suite (families below), formatted as chat with the gold answer in the final turn; source holds the benchmark name (belebele rows: facebook/belebele/<lang>).
  • ct / rl / persona .benchmark = held-out rows in each subset's own format (the benchmark suite has no chosen/rejected, verifier, or persona shape).

Validation + test + benchmark combined vs train: sft 35.4% · ct 33.5% · rl 33.4% · persona 33.3% — around the designed 1-per-3 ratio (v2's extra eval rows push sft slightly above; new training rows were split 75/12.5/12.5 to compensate).

Quickstart

from datasets import load_dataset

REPO = "MC7ever/all-in-one-llm-corpus"

sft      = load_dataset(REPO, "sft")       # chat SFT
dpo      = load_dataset(REPO, "ct")        # preference pairs
grpo     = load_dataset(REPO, "rl")        # RL tasks
persona  = load_dataset(REPO, "persona")   # persona SFT

# a single split of a single subset
train = load_dataset(REPO, "sft", split="train")

# language filtering (v2: every row carries a `language` column)
de = train.filter(lambda r: r["language"] == "de")
human_de = de  # all de rows are human-written (OASST1/Aya), slop-filtered

Suggested recipe

1. SFT — fit sft/train, pick checkpoints on sft/validation, report sft/test (in-distribution) and sft/benchmark (external).
2. DPO — start from the SFT model, fit ct/train, early-stop on ct/validation preference accuracy, report ct/benchmark.
3. GRPO — fit rl/train with the provided verifiers, monitor reward on rl/validation, report rl/benchmark.
4. Persona — mix persona/train into SFT or train a style LoRA; check adherence on persona/test.

# RL with verifiable rewards only (drop free-form tasks for strict scoring):
rl_v = load_dataset(REPO, "rl", split="train").filter(lambda x: x["verifier"] != "free")

Column reference

• language (v2, all subsets & splits) — ISO code from langdetect (en, de, uk, …) or unknown for very short/numeric prompts.

sft.subset tags (train split)

value	rows	meaning
general	147,499	general instruction/chat data (incl. human-written de/uk additions)
bench	37,631	benchmark-style training items (upstream-decontaminated)
reasoning	27,993	reasoning / math / CoT
legacy	23,302	recovered from blend v1–v2, predate the tagging scheme
persona	18,276	persona-flavoured chat
code	16,790	code
benchmark	55,867	(benchmark split only) external eval items

rl.verifier (full task set)

free 17,394 · persona 5,868 · number 2,194 · format 1,530 · span 1,217 · mcq 1,216 · python 1,130 · label 765 · sql 748 · yn 625 · json 432 · refusal_benign 378 · tool_outcome 378 · refusal_harmful 366

Benchmark families (sft.benchmark)

hellaswag 5,000 · mmlu_prox_ja 5,000 · bbh 4,999 · mmlu_prox_ko 4,999 · mmlu_prox_ru 4,999 · mmlu_prox_zh 4,999 · mmlu 4,995 · boolq 3,270 · squadv2 2,498 · arc_easy 2,376 · bb_emojis_emotion_mc 2,000 · piqa 1,838 · winogrande 1,267 · arc_challenge 1,172 · sciq 1,000 · belebele_de 900 · belebele_cs 900 · belebele_sk 900 · belebele_uk 900 · truthfulqa_mc1_mc2 805 · logiqa 651 · gpqa_diamond_zeroshot 198 · eq_bench 171 · aime24 30

Item kinds: mcq, free, yn, span, label, set, number.

🌍 Languages

langdetect over every row's first user prompt (language column, full detail in repo):

lang	rows	share
en	429,424	89.5%
ru	7,258	1.5%
ko	6,474	1.3%
zh	6,353	1.3%
unknown	6,335	1.3%
ja	6,252	1.3%
de	3,623	0.8%
uk	1,744	0.4%
es	1,510	0.3%
fr	1,139	0.2%
cs	1,056	0.2%
sk	974	0.2%
ro	860	0.2%
ca	758	0.2%

Coverage intent: en, ja, de, zh, sk, cs, ru, uk (+ existing ko and a ~1% long tail, kept per policy — the language column makes any subset trivial to select).

🧹 Cleaning & provenance

Built from 10 source datasets by deterministic script (build.py), then the v2 pass (build_v2.py):

family	sources	what happened
chat blends	super-lfm-blend-v1…v4 (525,109 rows)	concatenated, SHA1-deduplicated on normalized message content — 203,324 cross-version duplicates removed, newest version wins; 321,785 unique remain
RL tasks	super-lfm-rl-tasks-v1, v2	identical task IDs in both → v2 kept (richer metadata)
preference pairs	super-lfm-pref-v1, v2	identical content in both → v2 kept
benchmarks	super-lfm-eval-v1	9 exact duplicates removed → 52,268; reformatted to chat; lives only in sft.benchmark
persona	super-lfm-persona-v1	161 exact duplicates removed → 4,228
human languages (v2)	OpenAssistant/oasst1 (de/cs/sk/uk), CohereLabs/aya_dataset (de/uk)	2,276 + 760 = 3,036 rows accepted after dedup + the same slop filter, split 75/12.5/12.5 into train/validation/test
reading benchmarks (v2)	facebook/belebele (deu_Latn, ces_Latn, slk_Latn, ukr_Cyrl)	3,600 rows appended to sft.benchmark only

• Every assistant turn carries the <think></think> think-fix prefix (already present across all blend versions).
• Splits use a fixed seed (42) shuffle with exact-count slicing — rebuilds reproduce bit-for-bit.
• Upstream blends report decontamination against 20 common benchmarks; independently, this repo guarantees benchmark rows never enter a training split.

v2 de-bot pass — rows removed

rule	rows removed
bot-identity / corpo-speak phrase in the target text	10,665
same rules hitting DPO chosen (rejected kept as-is)	5,787
purple-prose density (heretic noslop markers ≥3/100 words)	2,334
same rules hitting RL gold answers	1,054
malformed conversation shape	23
AI-identity probe prompts ("are you an AI?")	11
total	19,874

Design notes:

• ct.rejected is never removed for slop — dispreferred bot-speak is exactly the negative signal DPO needs. Only chosen (and assistant/gold targets everywhere else) is filtered.
• persona content is untouched (measured 0% slop; it's deliberate creative voice).
• Purple-prose is removed by density, not single words — human horror/creative writing uses words like velvet and crimson too; AI slop clusters them (≥3 markers per 100 words, or ≥5 distinct in one target).
• Assistant-identity system lines ("You are a helpful assistant.") are stripped rather than dropping the row.

📊 Results & scores

Filter quality (measured)

Scored against DrRiceIO7/SlopReview v3 — 6,579 labeled AI-slop rows and 1,000 labeled human rows:

metric	score
AI-slop rows caught by the v2 filter	56.6%
human rows wrongly removed (false-positive rate)	8.2%
pre-v2 slop prevalence (phrase scan)	sft 3.2% · ct 19.8% · rl 2.1% · persona 0.0%

Slop inside ct.rejected is counted as detected here but deliberately kept in the dataset (negative examples).

Reference local run (fill-in)

Designed for a small model on a 16 GB machine — MLX (mlx-lm) or Unsloth, LION-style optimizer:

model	stage	steps / epochs	tokens	sft.test	sft.benchmark	ct.benchmark pref-acc	rl.benchmark reward
LFM 2.5 2.6B	SFT→DPO→GRPO	—	—	—	—	—	—
Spark 2.6 4B	SFT→DPO→GRPO	—	—	—	—	—	—

Eval protocol (same for any model you put in the table):

1. sft.benchmark — exact-match / letter accuracy per family (mcq, yn, span, number, set, label).
2. ct.benchmark — preference accuracy: share of pairs where the model ranks chosen > rejected.
3. rl.benchmark — mean verifier reward, macro-averaged over families (exclude free for strict scoring).
4. persona.benchmark — held-out persona adherence (style rubric / judge score).
5. Report the same metrics on test splits as the in-distribution counterpart.

⚠️ Content notice

Inherited from source datasets: the chat blends contain explicit horror and NSFW creative writing plus material from public roleplay and "uncensored" community datasets. Published for research and private experimentation; not filtered for workplace or educational use.

License

other — mixed upstream licenses; see the source dataset cards.