18 years of real production traffic from one of the longest-running conversational AI platforms in the world. Not synthetic. Not crowdsourced. Not scraped. Real users having real conversations with chatbots, collected with consent under platform Terms of Service from 2008 through 2026.

Why this dataset is different:

Real human intent: Users typing to chatbots they chose to interact with, not paid annotators.
Exceptional depth: Averages over 60 turns per session in the primary companion corpus, with interactions extending to thousands of turns.
Production diversity: 7 channels, 20+ industry verticals, 100+ languages in the full corpus.
Genuine adversarial inputs: Real users probing limits, not red-teamers simulating.
Clean provenance: Consent secured, copyright cleared, PII removed, GDPR/EU AI Act compatible.
Enterprise governance: Reviewed by external legal counsel as part of commercial licensing agreements with leading AI platforms.

Sample Configurations

This sample is split into 3 configurations. Load any one or combine.

from datasets import load_dataset

# Pick a configuration
ds = load_dataset("pandorabots/real-human-chatbot-dialogue-31B-sample", "companion_bot")
ds = load_dataset("pandorabots/real-human-chatbot-dialogue-31B-sample", "vertical_developer_bots")
ds = load_dataset("pandorabots/real-human-chatbot-dialogue-31B-sample", "multilingual")

Config	Convos	Turns	Languages	Best For
`companion_bot`	59	4,501	English	Long-context dialogue, persona modeling, safety/alignment
`vertical_developer_bots`	51	4,000	English, French, Spanish, Polish	Domain adaptation, conversational agents
`multilingual`	5	44	Arabic, Japanese, Hindi, Vietnamese, Korean	Multilingual demonstration & fallback tracking

Total: 115 conversations, 8,545 turns. Sample selected for quality; full corpus is vastly larger (see below).

Configuration Details

`companion_bot`

Long-running open-domain AI persona (Kuki / Mitsuku, multiple-time Loebner Prize winner) deployed across 7 channels: Discord, Facebook Messenger, Telegram, Roblox, Viber, web portal, and YouTube Live streaming.

Segment Highlights:

Extreme context length: Averages 75+ turns per session in this sample, with some extending past 800 turns. This is among the longest-context real human-AI conversational data publicly available.
Authentic emotional range: Joy, frustration, loneliness, humor, and curiosity expressed by real users.
Real adversarial inputs: Users testing limits, attempting jailbreaks, and probing for inconsistencies.
Persona consistency: A single named persona maintained across thousands of turns.
Multi-channel formats: Demonstrates how communication style shifts between gaming (Discord, Roblox), messaging (Telegram, Messenger), and broadcast (YouTube Live).

`vertical_developer_bots`

Real production chatbots built by independent developers and brands across 20+ industry verticals. Some are commercial enterprise deployments; others are indie projects. Includes both paid-tier and free-tier service samples for transparency.

Verticals represented:

Vertical	Description
Fashion retail	Shopping assistants and product quizzes
Apparel retail	Fit advice and customer support
Automotive aftermarket	Customer service for tire dealerships
Legal	Support and guidance bots
Online gambling	Sports betting support
Financial / credit	Credit scoring assistance
Consumer health and beauty	Personal care advisors
E-commerce / pharmacy	Order and product support
Language learning	English and bilingual practice apps
Employment	Career guidance and job search
Enterprise service desk	Internal IT and operations
Social issues	Discussion bots on contemporary topics
Consumer IoT	Smart device personas
Encrypted messaging	Native platform assistants
Brand training	Internal training for brand ambassadors
B2B manufacturing	Service provider chatbots
SaaS	Customer support bots
CRM	Platform support bots
Media	Music and entertainment chatbots
Mobile assistants	Early-era voice assistant interactions

`multilingual`

A curated demonstration sample highlighting both native-language interactions and the structural routing constraints of early-era English-based AIML architectures.

Code	Language	Convos	Notes
ar	Arabic	1	Educational bot answering physics questions natively.
ja	Japanese	1	Open domain interaction showing high native fidelity.
hi	Hindi	1	Demonstrates English framework fallback constraints.
vi	Vietnamese	1	Demonstrates English framework fallback constraints.
ko	Korean	1	Multi-turn open domain conversational sample.

This is a tiny cross-section, not a representative sample of the full multilingual corpus. The full corpus includes 100+ languages — see the Full Language Breakdown below.

About multilingual coverage in the broader corpus: The Pandorabots platform uses a wide variety of bots. Some bots (like the Arabic and Japanese samples) were built with localized native-language response logic. Many bots — including the primary Kuki/Mitsuku persona — are English-based AIML systems that have been deployed to users in many languages. For those deployments (like the Hindi, Vietnamese, and Korean samples), the unique value is the consent-cleared native-speaker INPUT distribution, which is scarce for many languages and useful for multilingual training even where bot output is not exemplary.

Samples in additional languages — including German, French, Spanish, Portuguese, Russian, Dutch, Italian, Korean, Japanese, Vietnamese, Mandarin, Hindi, and various low-resource languages — are available on request under NDA. Contact data@pandorabots.com.

Data Format

Each record in each .jsonl file is a single conversation:

{
  "conversation_id": "companion_kuki_hq_0001",
  "channel": "web_portal",
  "language": "en",
  "turns": [
    {"role": "user", "content": "I like your name"},
    {"role": "assistant", "content": "Yes Kuki is a very nice name isn't it?"}
  ]
}

Field	Description
`conversation_id`	Unique identifier
`channel`	`web_chat`, `web_portal`, `messaging`, `gaming`, `social_streaming`
`language`	ISO 639-1 code
`turns`	Ordered list of `{"role": "user" or "assistant", "content": "..."}`
`use_case`	(vertical only) Generic business use case description
`bot_id`	(vertical & multilingual only) Anonymized bot identifier
`service_tier`	(vertical only) `paid` or `free` — service tier of the developer's bot

PII & Redaction Conventions

[REDACTED] — placeholder for content that was masked at source (typically names, but in foreign-language conversations may include other identifying terms)
[IMAGE] — images that the bot sent inline
[LINK] — external links shared in conversation
[NAME REDACTED] — full names (where they appeared)
[EMAIL REDACTED] — email addresses
[BRAND REDACTED] — brand names that appeared in conversation content
First names alone are retained — they are not considered identifying PII under standard frameworks and are research-relevant for studying persona/name-following behavior

Limitations & Considerations

While this dataset offers rare, authentic human-bot interactions, researchers should be aware of several structural characteristics and artifacts inherited from a live production environment:

Variable Bot Quality & Fallback Loops: Unlike distilled LLM datasets (e.g., GPT-4 synthetic outputs), the assistant responses in this corpus come from deployed AIML (rules-based) systems and earlier-era intent classifiers. The bots frequently rely on scripted fallback responses (e.g., "I don't understand") and can occasionally get caught in repetitive conversational loops when users spam the system. The primary value of this dataset is the authentic human input distribution, not necessarily the bot's capabilities as a role model for generation.
Multimodal UI Artifacts & System Tokens: Because these bots were deployed on rich graphical interfaces (Messenger, Discord, web portals), user actions like clicking buttons or selecting carousels are logged as raw system tokens (e.g., XREACT <emotion>, XREADING <feature>, xpayload, endlesstrivia). These tokens are preserved because they encode meaningful multimodal interaction signals, but they will require filtering or special tokenization depending on your downstream training objective.
Unfiltered Human Toxicity: Open-domain, anonymous chatbots naturally attract boundary-testing. The dataset contains genuine instances of toxic language, explicit roleplay attempts, and aggressive adversarial probing. While highly valuable for RLHF and safety alignment, researchers should apply appropriate content warnings when distributing derivatives.
Mismatched Multilingual Routing: Outside of specifically localized bots, many non-English user inputs in the broader corpus were routed to English-centric AIML architectures. While this provides a massive, consent-cleared source of native-speaker input text, the corresponding bot outputs in those specific interactions are often structurally limited.

Provenance & Compliance


✅ GDPR / EU AI Act compatible	Reviewed by external legal counsel
✅ User consent secured	Via platform Terms of Service accepted at registration
✅ Copyright cleared	Platform-owned content under ToS
✅ PII removed	Names, usernames, URLs, emails redacted
❌ No synthetic data	All real human inputs
❌ No web scraping	Platform-native collection only

Intended Uses

Use Case	Why This Data Helps
LLM grounding & fine-tuning	Naturalistic, unscripted human dialogue at scale
RLHF / preference modeling	Long multi-turn context with real user signals
Safety & alignment training	Real adversarial and guardrail-probing inputs
Long-context evaluation	Avg 70 turns — genuine persistent context
Companion AI training	Persona consistency across thousands of turns
Domain adaptation	20+ real verticals with production task flows
Multilingual fine-tuning	5 non-English languages in sample, 100+ in full corpus
Low-resource language research	Consent-cleared data in scarce-for-AI languages
Conversational agent training	Multi-step task-oriented dialogue across industries
Dialogue coherence research	Unusually long sessions reveal turn-level coherence patterns

Commercial use of this sample, or use of the full corpus in any form, requires a separate license. Contact data@pandorabots.com.

Full Corpus

This sample represents a tiny fraction of the full corpus — well under 0.001%.

Segment	Volume
Companion Bot	~12.4B tokens
Vertical Developer Bots	~18.6B tokens
Multilingual Subset (100+ languages)	~1.5B tokens
Full Corpus	31B+ tokens

Roughly 10B human input tokens. Average 61 turns per session in the full Companion Bot segment, with sessions extending up to thousands of turns.

Full Language Breakdown (Top Languages by Token Volume)

Language	Tokens
English	12.5B
German	542.8M
Spanish	194.9M
Portuguese	177.9M
French	131.6M
Russian	104.0M
Dutch	101.7M
Italian	47.3M
Chinese	21.0M
Polish	20.2M
Turkish	15.4M
Hungarian	10.0M
Swahili	9.0M
Tagalog	8.6M
Arabic	7.9M
Korean	6.3M
Japanese	6.1M
Persian	4.5M

Plus 80+ additional languages including low-resource languages scarce for AI training. Full breakdown available on request.

Available Tranches

The full corpus can be licensed in segments or as a whole. Common configurations:

Companion Bot only (~12.4B tokens) — ideal for safety/alignment, persona, and long-context research
Vertical Developer Bots (~18.6B tokens) — ideal for domain adaptation and conversational agent training
Multilingual Subset (~1.5B tokens) — licensable as a full bundle or by individual language cluster
Full Corpus (31B+ tokens) — volume discounts available

Flexible terms — segment, language cluster, or full corpus licensing all available.

→ data@pandorabots.com

Citation

@dataset{pandorabots_real_human_chatbot_dialogue_31b_sample_2026,
  title   = {Pandorabots Real Human Chatbot Dialogue 31B (Evaluation Sample)},
  author  = {Pandorabots, Inc.},
  year    = {2026},
  url     = {https://huggingface.co/datasets/pandorabots/real-human-chatbot-dialogue-31B-sample},
  license = {CC BY-NC 4.0}
}

License

CC BY-NC 4.0 — free for research and non-commercial use with attribution. Commercial use requires a license agreement.

Pandorabots, Inc. — pandorabots.com — data@pandorabots.com

Downloads last month: 46

Total file size:

677 kB