local-model-explorer-data

An anonymous record of what people try to run locally, gathered by Local Model Explorer: the hardware they plan for (GPU memory, number of cards, system or unified memory), the models and context lengths they look at, which GGUF quants they open and copy commands for, and the llama-bench results and reports they choose to share.

The question it answers: what hardware do local LLM users have, what do they try to run on it, and how fast does it actually go?

Quick start

from datasets import load_dataset

df = load_dataset("LocalLLaMA/local-model-explorer-data", "events", split="train").to_pandas()
df = df[df.suspicious_flags.apply(len) == 0]                      # clean rows only

# One row per visit: keep the newest summary of each session
visits = (df[df.event_type == "session"].sort_values("timestamp")
            .drop_duplicates("session_id", keep="last"))
visits.vram_class.value_counts()

# Community llama-bench results: generation speed per model, quant and GPU
bench = df[df.event_type == "llama_bench"]
bench.groupby(["selected_model", "quant", "gpu_name"]).generation_tps.median()

What is recorded

The dataset records outcomes, not clicks.

Schema (schema_version 1)

Every row has every column; columns that don't apply to an event type are null.

suspicious_flags

Flagged rows are kept so filters stay transparent; the Space leaves them out of its statistics.

Layout and updates

data/incoming/YYYY/MM/DD/<instance>-<UTC time>-<suffix>.parquet   # written by the Space, at most hourly
data/events/YYYY-MM.parquet                                       # one file per finished month

Early each month the previous month is compacted into data/events/YYYY-MM.parquet (re-sent session rows reduced to the newest per visit) and its incoming files are deleted in the same commit. The events config reads both locations.

Privacy

Never collected: accounts, names, emails, IP addresses, cookies, user agents, device fingerprints, or individual clicks. Pasted llama-bench output is parsed on the server and discarded; only the numbers and settings above are stored, never file paths, CPU model strings or build machines. Hardware is stored as classes. Browsers sending Global Privacy Control or Do Not Track send nothing unless the user opts in.

Limitations

• Self-selected. Visitors are not a random sample of local LLM users.
• Planned, not owned. Hardware on session rows is what a visitor planned for, which may not be what they have. Benchmarks and reports are more reliable.
• Self-reported. Benchmark output can be edited before pasting; implausible values are flagged, not proven.

License

CC BY 4.0.