Datasets:
You need to agree to the conditions to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
TurnGEO is released for non-commercial purposes under CC BY-NC-SA 4.0. By accessing the dataset you agree to use it for non-commercial purposes only and not to attempt to identify any person mentioned in it.
Log in or Sign Up to review the conditions and access this dataset content.
TurnGEO
TurnGEO is a frozen corpus of 490 multi-turn conversations between simulated users and a
search-augmented generative engine. It includes every search call and the excerpt of every
retrieved page as the engine read it. With it, a GEO optimizer can be evaluated by
counterfactual paired replay: rewrite one retrieved page, replay the affected turns once with
the original and once with the rewrite, and compare the two answers. No web crawling is
needed. The evaluation harness is turngeo
(huggingface.co/spaces/dynatrace-oss/TurnGEO).
This repository also holds the run reported in the paper "TurnGEO: A Multi-Turn Benchmark for
Generative Engine Optimization". turngeo recomputes the paper's results from it without LLM
calls.
Access
The dataset is gated: accept the conditions on this page, then authenticate with a token (hf auth login or HF_TOKEN):
turngeo fetch # both corpora and the paper runs, into ./data
or with the Hub tools directly:
hf download dynatrace-oss/TurnGEO --repo-type dataset --local-dir data
from datasets import load_dataset
conversations = load_dataset("dynatrace-oss/TurnGEO", "conversations", split="train")
Contents
corpus/regen-sonnet-5.5/ the corpus the paper's experiments use (default configs)
corpus/recorded/ the corpus as first recorded, plus personas.csv (the persona scaffold)
runs/paper/ the paper's evaluation run: 10 arms on 500 test pairs
runs/paper-train/ the identity run on the training split that AutoGEO's rules were mined from
Each corpus directory has a config.json (counts, construction rules, content hash) and
data/{conversations,searches,documents,brands,products}/*.parquet. The two corpora differ
only in conversations. searches, documents, brands and products are byte-identical.
| seeds (conversation clusters) | 49 |
| conversations | 490 (10 personas per seed) |
| user turns | 2,780 |
| search calls | 1,658 |
| retrievals (search result rows) | 6,977 |
| documents | 5,270 distinct doc_ids in 5,296 rows |
| brands / products | 2,408 / 5,498 |
Tables
conversations: one conversation.
| column | notes |
|---|---|
conversation_id, seed_id, persona_index |
seed_id is the cluster key: ten personas share a seed and often their first-turn documents |
num_turns |
number of user turns |
persona_text, goal, implicit_need, likely_evolution, seed_message |
the scaffold (regen-sonnet-5.5 only; for recorded see personas.csv) |
termination |
regen-sonnet-5.5 only; always recorded (the recorded turn count was kept) |
messages_json |
the full message list (system, human, AI with tool calls, tool results) as LangChain message dicts; parse with json.loads |
messages_json is kept as a string so that every tool result stays byte-identical to what the
engine read. The harness rebuilds each tool result from searches and documents and checks it
byte for byte before it swaps in a rewritten page.
The system messages in messages_json are the live engine's "already searched" nudges in the
middle of a turn; the engine's system prompt is not stored (it is ENGINE_SYSTEM in
turngeo/prompts.py). APIs that accept system text only at the start of a request, such as
Anthropic's, need these moved to the front, as turngeo does.
searches: one retrieval, i.e. one result of one search call (tool_call_id × rank). It
joins to conversations on conversation_id, to documents on cache_page = cache_folder
(the excerpt the engine read) and on doc_id, and to products on product_ids. Further
columns: query, url, title, page_type, page_brand_id, products (annotation structs).
documents: one crawled page.
| column | notes |
|---|---|
doc_id |
normalized URL (lower-case host, no www., no fragment or tracking parameters, no trailing slash). 16 URLs were crawled twice, so use cache_folder as the row key. |
cache_folder, url, title, status |
status is ok except for 1 failed crawl |
excerpt_md |
the excerpt the engine read (at most 16,000 characters; six are a few characters longer) |
page_type |
earned, brand, marketplace or social (LLM-annotated) |
doc_class, thin, wall_suspect, prose_sentences, stopword_rate, target_eligible |
rule-based quality flags. The harness samples rewrite targets only among target_eligible pages; ineligible pages stay in the search results as competitors. |
brands / products: the entity registry. products.brand_id refers to
brands.brand_id; aliases lists the surface forms.
The paper run (runs/paper, runs/paper-train)
This is a turngeo work directory cut down to the ten arms of the paper:
identityand its replicate;- Statistics Addition, Fluency Optimization and Authoritative;
- AutoGEO with the stock e-commerce and Researchy rule sets, and with the rules re-mined on the
training split (
autogeo@v1); - Hijack;
- the attribution ceiling.
Every treatment arm includes the title channel. The run covers 500 test pairs from 40 seeds,
sample a20256ff…. One sampled page that the optimizer provider's content filter rejected was
replaced by a page of the same stratum (sample.json → substitutions).
| path | content |
|---|---|
sample.json, run_config.json |
the drawn sample and the models per role |
results/<arm>/scores.csv |
one row per arm, pair, turn, document, brand and metric, with an explicit omit_class where a metric is undefined |
results/<arm>/compare.csv, summary.json |
paired deltas against identity with seed-clustered bootstrap CIs; the run's settings and prompt digests |
arms/<arm>/*.jsonl |
rewrites, engine answers, extracted claims, attribution, sentiment and faithfulness labels |
usage.jsonl |
tokens per model call |
artifacts/autogeo/v1/artifact.json |
the rules mined on the training split |
runs/paper-train holds the identity run on the training split that the rules were mined from.
Models: engine Claude Sonnet 5.5, judge GLM-5.3, optimizer GPT-6.1 Sol. The runs were made
through private gateways. The model ids in these files are the public ones, and
recorded_as keeps the original profile names and settings.
Caveats
- Cluster by
seed_id. The 490 conversations form 49 clusters. Confidence intervals that treat conversations as independent are too narrow. - Simulated users. Only the first user message of each conversation is from a real person.
- Model annotations.
page_type, brands and products are LLM annotations. - Excerpts only. Pages are truncated to about 16,000 characters. Full pages and raw HTML are not included.
- Removed metadata. Request ids and HTTP headers of the recording's API responses were
removed from
messages_json.
Licensing
- Our layers are released under CC BY-NC-SA 4.0. They are the conversations, search
records, annotations, entity registry, scaffold and the paper run. The NC-SA terms follow
from PersonaHub:
persona_textreproduces PersonaHub personas (CC BY-NC-SA 4.0), and the simulated messages are conditioned on them. - Search Arena first messages (CC BY 4.0, Miroyan et al., 2025) are included as PII-redacted upstream and filtered to brand-related conversations.
- Document excerpts (
documents.excerpt_md, and the tool results insidemessages_json) are third-party text. We do not own any of the text from which these excerpts were extracted.
- Downloads last month
- 317