OpenChatData / README.md
gptforfree's picture
Update README.md
eeb4e69 verified
|
Raw History Blame Contribute Delete
9.8 kB
metadata
license: apache-2.0

OpenChatData

OpenChatData is an anonymized dataset derived from database dumps from a discontinued AI chatbot service that routed model requests through OpenRouter.

The dataset contains 20,949 chat-log records collected between February 4, 2026 and April 5, 2026, covering usage across 27 model identifiers.

Important: OpenChatData does not contain the text of user prompts or model responses. The released data consists of metadata and aggregate measurements such as token, word, and character counts.

Dataset Overview

Each row represents a recorded chatbot interaction/request.

The dataset includes:

  • Date of the interaction
  • Coarse geographic information
  • OpenRouter model identifier
  • Prompt/query token counts
  • Model response token counts
  • Word counts
  • Character counts
  • Whether the record was marked hidden by the original service

The supplied dataset contains 15 columns and 20,949 rows.

Schema

Column Description
date Date associated with the interaction
city City associated with the request
state State/region associated with the request
country Country code associated with the request
openrouter_model_slug Model identifier used for the request
query_tokens Number of tokens in the user's query
response_tokens Number of tokens in the model response
total_tokens Combined query and response token count
query_words Number of words in the user's query
response_words Number of words in the model response
total_words Combined query and response word count
query_chars Number of characters in the user's query
response_chars Number of characters in the model response
total_chars Combined query and response character count
is_hidden Integer flag indicating whether the interaction was marked hidden in the original service

Basic Statistics

Property Value
Records 20,949
Columns 15
Date range 2026-02-04 – 2026-04-05
Unique model identifiers 27
Unique countries 5
Unique cities 5
Records marked hidden 18,684

No missing values are present in the 15 released columns.

Models

Requests in the dataset were routed through a variety of models available through OpenRouter. Model names are preserved as they appeared in the original database.

As a result, model identifiers are not necessarily normalized. For example, identifiers such as:

gpt-4.1
openai/gpt-4.1

may refer to the same underlying model while appearing as separate values in the dataset.

Some of the most frequently occurring identifiers include:

Model identifier Records
gpt-4.1 4,717
openai/gpt-4o 3,414
openai/gpt-3.5-turbo 2,472
openai/gpt-4.1 2,152
openai/gpt-oss-120b 2,007
openai/gpt-oss-120b:free 1,268
qwen/qwen-max 1,068
nvidia/nemotron-nano-9b-v2:free 1,001

Researchers may wish to normalize model identifiers before performing model-level comparisons.

Privacy and Anonymization

OpenChatData originates from database dumps belonging to an abandoned chatbot service.

The public dataset intentionally excludes the actual contents of conversations. In particular, it does not include:

  • User prompts
  • Model response text
  • Usernames or display names
  • Email addresses
  • Passwords or authentication credentials
  • IP addresses
  • Session or conversation contents

The remaining geographic fields should still be treated as potentially sensitive metadata. Users of the dataset should avoid attempts to identify, profile, or deanonymize individuals represented by the data.

The is_hidden field is retained as metadata from the original service. A hidden record should not be interpreted as permission to reconstruct, infer, or expose the underlying conversation.

Potential Uses

OpenChatData may be useful for research and analysis involving:

  • LLM usage patterns
  • Model popularity
  • Prompt and response length distributions
  • Token consumption
  • Differences in response length between models
  • Usage patterns over time
  • Model-routing analysis
  • OpenRouter ecosystem analysis
  • Synthetic workload generation
  • LLM infrastructure and capacity-planning research

Because conversation text is not included, the dataset is better suited to usage and workload analysis than semantic analysis of chatbot conversations.

Example

import pandas as pd

df = pd.read_csv("openrouter_chat_data.csv")

print(df.shape)
print(df.columns)

# Most frequently used models
print(df["openrouter_model_slug"].value_counts().head(10))

# Average token usage by model
model_stats = (
    df.groupby("openrouter_model_slug")["total_tokens"]
      .agg(["count", "mean", "median"])
      .sort_values("count", ascending=False)
)

print(model_stats)

Data Provenance

The data was recovered from database dumps from a now-discontinued AI chatbot service.

The service used OpenRouter to provide access to multiple language models. OpenChatData is an archival release of anonymized metadata extracted from those records rather than an official OpenRouter dataset.

OpenChatData is not affiliated with, endorsed by, or produced by OpenRouter or the model providers represented in the dataset.

Model identifiers and provider names remain in the dataset because they describe which models handled requests at the time the service was operating.

Limitations

There are several important limitations when interpreting this dataset.

Sampling bias. The records represent users of one chatbot service and should not be treated as representative of OpenRouter traffic, global LLM usage, or the general population.

Model-name inconsistencies. Model identifiers were stored by the original application and may contain aliases, prefixes, or naming conventions that changed over time.

Geographic precision. Location fields reflect information stored by the original service and may be inaccurate, approximate, VPN-influenced, or otherwise unrelated to a user's actual location.

No conversation semantics. Since prompt and response contents have been removed, the dataset cannot directly answer questions about conversation topics, intent, response quality, safety, or sentiment.

Historical snapshot. The dataset reflects a limited period of activity from February through April 2026 and should not be interpreted as describing current model availability or popularity.

Application-specific behavior. Token counts, hidden flags, routing decisions, and other fields may reflect implementation details of the original chatbot service.

Responsible Use

This dataset is intended for research, education, statistical analysis, and other legitimate uses.

Do not use OpenChatData to attempt to identify or deanonymize former users of the service. Researchers publishing analyses based on the dataset should consider aggregating geographic and low-frequency categories where necessary to avoid exposing potentially identifying usage patterns.

Citation

If you use OpenChatData in research, please cite the dataset and its repository.

A formal citation can be added here once a DOI or permanent repository URL is available.

@dataset{openchatdata2026,
  title  = {OpenChatData},
  year   = {2026},
  note   = {An anonymized dataset of OpenRouter-backed chatbot usage metadata}
}

License

OpenChatData is licensed under the Apache License 2.0.

You may use, modify, and redistribute the dataset in accordance with the terms of the Apache License, Version 2.0. See the repository's LICENSE file for the full license text.

Unless required by applicable law or agreed to in writing, the dataset is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.

Disclaimer

OpenChatData is an independent archival dataset derived from a discontinued third-party chatbot service.

OpenRouter and the individual model providers represented by model identifiers in the dataset are not responsible for the collection, preparation, anonymization, publication, or maintenance of this dataset.