KapInstruct-100M: Curated 100-Million Token Instruction Tuning Dataset

KapInstruct-100M is a high-fidelity, 100-million-token instruction-tuning dataset engineered for Supervised Fine-Tuning (SFT) and alignment of compact language models (under 1 billion parameters). Formatted with the Qwen ChatML chat template and tokenized using Qwen/Qwen3.5-0.8B-Base, the dataset enforces strict assistant-only loss masking (masking user prompts and structural delimiters to -100) to maximize training efficiency.

KapInstruct-100M unifies 12 balanced, high-signal instruction sources spanning programming synthesis, step-by-step mathematical reasoning (Chain-of-Thought), technical STEM QA, multi-turn dialogue, strict constraint following, and interactive code debugging/repair.

---

Dataset Overview

• Hugging Face Repository: kaptaan45/KapInstruct-100M
• Kaggle Dataset: kaptaan45/kapinstruct-100m
• Total Usable Tokens: 100,000,000 tokens post-filtering, normalization, and deduplication
• Packed Sequence Length: 4096 tokens per packed sequence
• Tokenizer: Qwen/Qwen3.5-0.8B-Base (248,044 BPE vocabulary)
• Loss Masking Policy: assistant_only (prompts, system messages, and <|im_start|> headers have labels = -100; loss is computed strictly on assistant response spans)
• Primary Storage Formats: Memory-mapped Apache Arrow IPC (.arrow) and Apache Parquet (.parquet)
• Primary Use Case: SFT / Instruction Tuning for compact code and reasoning models such as QaptaanLM-0.75B.

---

Motivation & Design Principles

Supervised fine-tuning of compact models (0.5B to 1.5B parameters) is highly sensitive to data quality and token loss allocation:

1. Assistant-Only Loss Masking: Standard causal LM training over unmasked instruction data wastes gradient updates predicting user prompts and static system headers. By masking all non-assistant tokens to -100, 100% of gradient updates focus on assistant reasoning, syntax accuracy, and answer generation.
2. Deficit-Weighted Balanced Scheduling: Rather than concatenating disparate dumps, KapInstruct-100M uses a deficit-driven sampling scheduler that measures exact tokenizer tokens post-filtering, guaranteeing precise representation across all 12 domains.
3. Cross-Source Global Deduplication: Full multi-turn dialogues are canonicalized and indexed via SHA-256 to eliminate prompt leaks, dataset overlaps, and synthetic duplicates across independent upstream sources.
4. Rich Reasoning Traces (Chain-of-Thought): Mathematics and STEM partitions retain detailed step-by-step reasoning solutions, empowering compact models to learn structured problem breakdown.

---

Dataset Composition & Source Mixture

KapInstruct-100M is composed of 12 verified upstream sources sampled according to strict token budgets:

Source	Domain / Category	Share	Tokens	License
Smol-Magpie-Ultra	General reasoning & conversation	18%	18,000,000	Apache-2.0
Magicoder-Evol	Complex programming instructions	13%	13,000,000	Apache-2.0
OpenMathInstruct-2	Math problem solving & synthesis	11%	11,000,000	CC-BY-4.0
CodeFeedback-Filtered	Bug fixing & code repair	10%	10,000,000	Apache-2.0
OpenHermes-2.5	Broad conversational QA	9%	9,000,000	MIT
Magicoder-OSS	Open-source code generation	8%	8,000,000	MIT
OpenThoughts-114k	General & STEM reasoning	7%	7,000,000	Apache-2.0
NuminaMath-CoT	Competition math reasoning	6%	6,000,000	Apache-2.0
Tulu-3 SFT	High-fidelity instruction following	6%	6,000,000	ODC-By
Self-OSS StarCoder2	Execution-validated code	5%	5,000,000	ODC-By
WebInstructSub	Science & technical QA	4%	4,000,000	Apache-2.0
Smol-Constraints	Strict constraint adherence	3%	3,000,000	Apache-2.0
Total		100%	100,000,000

---

Domain Allocation Breakdown

+-------------------------------------------------------------+
|             KapInstruct-100M Domain Allocation              |
+-------------------------------------------------------------+
|  [================]  Code Generation (31% - 31M tokens)     |
|  [==============]    General Reasoning (27% - 27M tokens)   |
|  [=========]         Mathematics CoT (17% - 17M tokens)     |
|  [======]            STEM QA & Science (11% - 11M tokens)   |
|  [=====]             Debugging & Repair (10% - 10M tokens)  |
|  [==]                Constraint Adherence (4% - 4M tokens)  |
+-------------------------------------------------------------+

---

ChatML Formatting & Loss Masking

Each conversation is formatted strictly following the Qwen ChatML schema:

<|im_start|>system
You are a helpful and harmless assistant.<|im_end|>
<|im_start|>user
Write a function in Python to compute the Levenshtein distance.<|im_end|>
<|im_start|>assistant
def levenshtein_distance(s1: str, s2: str) -> int:
    ...<|im_end|>

Token-Level Alignment & Masking Verification

Turn Component	Rendered Token Span	Loss Label (labels)	Masking Status
System Turn	`<	im_start	>system\n...<
User Turn	`<	im_start	>user\n...<
Assistant Header	`<	im_start	>assistant\n`
Assistant Content	`response text...<	im_end	>\n`
Sequence Padding	`<	endoftext	>` infilling

In multi-turn dialogues [User 1 -> Assistant 1 -> User 2 -> Assistant 2], loss is computed strictly across Assistant 1 and Assistant 2 response spans.

---

Quality Filtering & Deduplication

1. Natural Language Filtering: FastText language identification and English confidence scoring (min_confidence = 0.65), with automatic preservation of code-mixed technical dialogues.
2. Programming Language Normalization: Canonical alias mapping across 16 core languages (Python, TypeScript, JavaScript, C++, C, C#, Java, Rust, Go, Ruby, PHP, SQL, Shell, HTML, CSS, Dockerfile).
3. Secret & Key Stripping: Regex scanning and complete rejection of leaked API keys (OpenAI sk-, AWS AKIA, GitHub ghp_, Slack, JWTs, and private RSA/SSH keys).
4. LaTeX & Math Integrity: Rejection of unbalanced LaTeX delimiters ($$, \begin{...}) and OCR noise artifacts.
5. Prompt Injection & Repetition Removal: Scanning and removal of injection jailbreaks, infinite loops, and degenerative repetition.
6. Global Cross-Source Deduplication: Exact SHA-256 fingerprinting on normalized dialogue turns across all 12 constituent datasets.

---

Dataset Loading and Usage

1. Zero-Copy Memory-Mapped PyArrow Loading (Fastest)

import glob
import pyarrow as pa
from datasets import load_dataset

# Load Arrow shards directly
shard_files = sorted(glob.glob("data/kapinstruct/*.arrow"))
dataset = load_dataset("arrow", data_files=shard_files, split="train", keep_in_memory=False)

print(f"Total packed sequences: {len(dataset):,}")
sample = dataset[0]
print(f"Sequence length: {len(sample['input_ids'])} tokens")
print(f"Trainable tokens: {sum(1 for l in sample['labels'] if l != -100)}")

2. Hugging Face Datasets Streaming

from datasets import load_dataset

dataset = load_dataset("kaptaan45/KapInstruct-100M", split="train", streaming=True)
sample = next(iter(dataset))
print("Loaded sequence keys:", list(sample.keys()))

3. PyTorch Training Loop Integration

import torch
from torch.utils.data import DataLoader

def collate_fn(batch):
    return {
        "input_ids": torch.tensor([b["input_ids"] for b in batch], dtype=torch.long),
        "attention_mask": torch.tensor([b["attention_mask"] for b in batch], dtype=torch.long),
        "labels": torch.tensor([b["labels"] for b in batch], dtype=torch.long),
    }

loader = DataLoader(dataset, batch_size=8, shuffle=True, collate_fn=collate_fn)
for batch in loader:
    # Forward pass computes cross-entropy loss ONLY on labels != -100
    outputs = model(
        input_ids=batch["input_ids"],
        attention_mask=batch["attention_mask"],
        labels=batch["labels"]
    )
    loss = outputs.loss
    loss.backward()
    break

---

Licensing and Provenance

KapInstruct-100M is a curated composite dataset. Each constituent subset retains its upstream license terms as documented in licenses.json:

Subset / Source	Upstream License	Attribution & Commercial Use
smol_magpie_ultra, smol_constraints	Apache-2.0 / Open	HuggingFaceTB / SmolTalk
magicoder_evol, code_debugging	Apache-2.0	ISE UIUC / M-A-P
magicoder_oss, openhermes_2_5	MIT	ISE UIUC / Teknium
openmathinstruct2	CC-BY-4.0	NVIDIA Corporation
numinamath_cot, openthoughts_reasoning	Apache-2.0	AI-MO / Open-Thoughts
tulu3_sft, self_oss_starcoder2	ODC-By	Allen AI / BigCode Project
stem_qa (WebInstructSub)	Apache-2.0	TIGER-Lab

Users and researchers must comply with the individual licenses of each constituent source.

---

Citation

To cite the KapInstruct-100M dataset in research:

@misc{kapinstruct100m2026,
  title   = {{KapInstruct-100M}: A Curated 100-Million Token Multi-Source Instruction Tuning Dataset for Compact Models},
  author  = {Kaptaan, Rudy and Contributors},
  year    = {2026},
  publisher = {Hugging Face},
  url     = {https://huggingface.co/datasets/kaptaan45/KapInstruct-100M}
}

@misc{qaptaanlm2026,
  title   = {{QaptaanLM-0.75B}: Efficient Hybrid-Attention Foundation Language Model},
  author  = {Kaptaan, Rudy and Contributors},
  year    = {2026},
  publisher = {GitHub},
  url     = {https://github.com/rudy-07/QaptaanLM-0.75B}
}