Research datasets
Immutable, versioned releases of public The Chat Archive records. Records under incompatible licenses are released as separate partitions.
Releases
Quick start
# List releases and the external dataset catalog
curl https://thechatarchive.com/api/datasets
# Release manifest: partitions, licenses, files, checksums
curl https://thechatarchive.com/api/datasets/v0.1
# Python
import pandas as pd
df = pd.read_parquet("https://thechatarchive.com/api/datasets/v0.1/files/community/community-00000.parquet")
df["messages"] = df["messages"].map(json.loads)
# Search the live archive (cursor pagination)
curl "https://thechatarchive.com/api/conversations?q=linked+list&model=gpt-4o&limit=50"Record schema (v1.0)
Public exports use a dedicated schema. It never contains consent records, account identifiers, IP hashes, moderation notes or upstream location and header fields.
| conversation_id | Stable ULID (lowercase). Never reused. |
| schema_version | 1.0 |
| source_type | user_contributed | external_dataset |
| source_dataset | Slug of the external dataset, if imported. |
| source_record_id | Record id within the external dataset (e.g. WildChat conversation_hash). |
| title | "string" |
| provider | ["string","null"] |
| model | ["string","null"] |
| model_verification | self_reported = stated by the contributor, not independently verified. |
| language | ["string","null"] |
| created_at | When the conversation originally happened, if known. |
| published_at | ["string","null"] |
| topic_tags | "array" |
| turn_count | "integer" |
| messages | "array" |
| feedback | "object" |
| attribution | Contributor username, pseudonym, 'anonymous', or upstream dataset citation. |
| license | "string" |
| url | "string" |
Collection & filtering methodology
Records are conversations that (1) were contributed by signed-in users who explicitly consented to public redistribution, passed automated sensitive-information scanning, contributor review/redaction, and human moderation; or (2) were imported from external datasets whose licenses permit redistribution, filtered per the dataset feasibility report, with upstream location/IP/header fields dropped. Records removed or suppressed before the snapshot are excluded.
Known limitations
Self-selected, non-representative sample of AI usage. Model names for community contributions are self-reported unless marked verified. Automated PII detection is imperfect; some personal information may remain. Imported datasets carry their own collection biases and licenses. Responses to similar prompts were produced under different, uncontrolled contexts and must not be treated as a benchmark.
Removals are listed in each release’s changelog as counts and reason categories only. Records removed after a release stay in copies people already downloaded; we ask users of past releases to apply later removals. See the dataset policy.