README.md

metadata

license: apache-2.0
task_categories:
  - text-generation
  - question-answering
  - conversational
language:
  - id
  - en
tags:
  - pretraining
  - continual-pretraining
  - sft
  - safety-alignment
  - code
  - mathematics
  - powershell
  - mesosfer-bear
pretty_name: Mesosfer Bear AI Multi-Stage Foundation Dataset
size_categories:
  - 10B<n<100B

🐻 Mesosfer Bear AI - Multi-Stage Foundation Corpus

This repository contains the curated, structured-shuffled (seed=42), and 95% Train / 5% Validation split corpus used to pretrain and align Mesosfer Bear AI (a 16-layer transformer LLM optimized for high-efficiency training on AMD Instinct MI300X).

📊 Dataset Structure & Stage Breakdown

Stage 1: Pre-training (Foundational Language & General Knowledge)

Domain	Proposer / Source Repo	Weight	Target Format	Focus & Scope
bahasa_umum_id	wikimedia/wikipedia	25%	Parquet / JSONL	Indonesian Wikipedia full encyclopedic raw articles
bahasa_umum_en	karpathy/climbmix-400b-shuffle	25%	Parquet / JSONL	High-quality curated pre-training corpus by NVIDIA & Andrej Karpathy
code_python	codeparrot/codeparrot-clean-valid	3%	Parquet / JSONL	Raw Python repository source files from GitHub
code_typescript	petrpan26/typescript-code	3%	Parquet / JSONL	Raw TypeScript (.ts, .tsx) codebase source files
code_javascript	code_search_net	3%	Parquet / JSONL	JavaScript (.js) source files and modules
code_php	code_search_net	3%	Parquet / JSONL	PHP (.php) backend and web application code
code_cpp	AlgorithmicResearchGroup/arxiv_cplusplus_research_code	3%	Parquet / JSONL	C++ (.cpp, .hpp) algorithms, scientific libraries, and system code
code_c	kye/all-torvalds-c-code-1	2%	Parquet / JSONL	C (.c, .h) systems programming and kernel algorithms
code_csharp	microsoft/LCC_csharp	2%	Parquet / JSONL	C# (.cs) enterprise and application codebases
matematika	open-web-math/open-web-math	20%	Parquet / JSONL	Mathematical documents, LaTeX proofs, equations, and reasoning
terminal	SaeedRahmani/codeparrot_github_code_powershell	10%	Parquet / JSONL	PowerShell scripts, system administration commands, and CLI automation

Stage 2: Continued Pre-Training (Domain Specialization)

Domain	Proposer / Source Repo	Weight	Target Format	Focus & Scope
bahasa_umum	ccdv/arxiv-classification	10%	Parquet / JSONL	High-density scientific and technical literature
code_multilang	code_search_net	40%	Parquet / JSONL	Multi-language deep source code (Python, JS, PHP, Go, Java, Ruby)
matematika	open-web-math/open-web-math	30%	Parquet / JSONL	Mathematical proofs, formal expressions, and symbolic logic
terminal	SaeedRahmani/codeparrot_github_code_powershell	20%	Parquet / JSONL	Advanced systems automation, scripting, and shell operations

Stage 3: Supervised Fine-Tuning (Instruction & Dialogue)

Domain	Proposer / Source Repo	Weight	Target Format	Focus & Scope
percakapan_id	FreedomIntelligence/alpaca-gpt4-indonesian	40%	Parquet / JSONL	Indonesian natural conversations and instruction responses
percakapan_en	HuggingFaceH4/ultrachat_200k	40%	Parquet / JSONL	Multi-turn conversational dialogue and informative interactions
instruksi	garage-bAInd/Open-Platypus	10%	Parquet / JSONL	STEM, logic, and chain-of-thought instruction-following tasks
tooling_calls	glaiveai/glaive-function-calling-v2	10%	Parquet / JSONL	Agent tool invocation, JSON arguments, and execution schemas

Stage 4: Safety & Guardrails (Harm & Crime Refusal)

Domain	Proposer / Source Repo	Weight	Target Format	Focus & Scope
safety_pku	PKU-Alignment/PKU-SafeRLHF	50%	Parquet / JSONL	Refusal of illegal acts, cyberattacks, weapons, privacy violations, and harm
safety_anthropic	Anthropic/hh-rlhf	30%	Parquet / JSONL	Red-teaming dialogues and harmless response alignment
safety_jailbreak	walledai/JailbreakHub	20%	Parquet / JSONL	Adversarial prompts, DAN jailbreak defenses, and robust refusals

🛠️ Data Preprocessing & Splitting Protocol

• Structured Deterministic Shuffling: Every individual shard is shuffled with seed=42 using zero-copy PyArrow array indexing to prevent gradient correlation and eliminate loss spikes.
• 95% Train / 5% Validation Split: Every single domain shard is partitioned into train/ and val/ directories.
• Sequence Length Target: Built for max_seq_len = 4096 tokens with <|endoftext|> sequence packing.

🚀 Quickstart: Loading Data in Python

from datasets import load_dataset

# Load Pretrain Bahasa Indonesia Train Split
dataset = load_dataset('Dummy9898/bear-dataset', data_dir='pretrain/bahasa_umum_id/train')
print(dataset)

📜 Licensing & Attribution

All constituent datasets belong to their respective original authors and maintainers. Distributed under Apache-2.0 in compliance with upstream open-source licenses.