LocalLLaMA/local-model-explorer-data

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.

event_type	One row per	Contents
session	visit	Hardware class, the filters the visit ended on, the top result and how it fit, and which models were opened, which repos were followed to the Hub and which run commands were copied. Re-sent as the visit changes: keep the newest row per session_id.
feedback	report	"Have you run this model?" with the quant, runtime, placement and how it went
llama_bench	benchmark	Numbers from a pasted llama-bench run, parsed by the Space

Schema (schema_version 1)

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

Column	Type	Rows	Description
timestamp	string	all	UTC, second precision
schema_version, app_version	int, string	all
event_type, session_id	string	all	session_id is random per browser tab and discarded when it closes
hardware_kind	string	all	gpu, unified, cpu
gpu_model	string	all	A listed card (e.g. NVIDIA RTX 3090) or custom
gpu_count	int	all	Cards
vram_class	int	all	Total GPU memory rounded down to a class: 0, 4, 6, 8, 10, 12, 16, 20, 24, 32, 40, 48, 64, 80, 96, 128, 160, 192, 256, 320, 384, 512 GB
ram_class	int	all	System (or unified) memory rounded down to a class
unified_system	string	all	Apple Silicon Mac, AMD Ryzen AI Max+ 395 (Strix Halo), NVIDIA DGX Spark
hardware_source	string	all	preset or custom
webgpu_vendor	string	session	WebGPU adapter vendor, e.g. apple, nvidia
browser_family, os_family	string	session	Coarse families
model_family, size_bucket	string	session, …	Final filters; on reports, the model's
context, kv_type	int, string	session	Planned context length and KV cache precision (f16, q8_0, q4_0)
priority, sort, fits_only	string, string, bool	session
result_count, families_searched, distinct_queries, searched	int, list, int, bool	session	What the visit looked at
top_model, top_quant, top_fit	string	session	First result and its fit: gpu, gpu_moe (experts in RAM), partial, cpu, unified, unified_tight, no
engine_version	string	session	Ranking engine, e.g. fit-v1
models_viewed, repos_clicked, commands_copied	list	session	Opened models (max 10), repos followed (max 5), llama.cpp / ollama
selected_model, repo, quant	string	reports	The model (base repo id, or a name for variants), the GGUF repo and quant
quant_bucket, params_b, moe, arch, repo_downloads_at_selection		reports	Filled in by the Space from the catalogue
tried, outcome, quality_rating, failure_reason	string	feedback	outcome is worked or problem
runtime, placement	string	feedback	llama.cpp, ollama, lm-studio, …; gpu, experts_in_ram, partial, cpu, unified
reported_tps, reported_context, notes	float, int, string	feedback	Notes ≤ 280 characters; emails, links and phone numbers are removed
prompt_tokens, prompt_tps, generation_tokens, generation_tps	int, float	llama_bench	From the pp and tg rows
backend, n_gpu_layers, n_threads, flash_attn, type_k		llama_bench	Settings reported by llama-bench
bench_model_type, bench_size_gib, bench_params_b	string, float, float	llama_bench	What llama-bench says it loaded
gpu_name, build	string	llama_bench	Generic GPU name parsed from the output (e.g. NVIDIA GeForce RTX 4090) and llama.cpp build commit
suspicious_flags	list	all	Empty when the row passed every check

suspicious_flags

Flag	Meaning
unknown_model, unknown_repo, unknown_model_in_session	Not in the catalogue at the time
repo_not_in_model	The repo holds a different model
implausible_tps	Faster than memory bandwidth allows for that model (with 3× headroom)
benchmark_model_mismatch	llama-bench's parameter count doesn't match the chosen model
unparseable_benchmark, incomplete_benchmark, incomplete_feedback	Missing numbers or the model
conflicting_feedback, implausible_hardware	Contradictory or impossible values

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.