README.md

---
pretty_name: "MLX Benchmarks"
license: apache-2.0
language:
- en
tags:
- benchmark
- evaluation
- llm
- mlx
- apple-silicon
- throughput
- latency
- code-generation
- reasoning
- math
size_categories:
- "n<1K"
configs:
- config_name: default
  data_files:
  - split: train
    path: "data/*.parquet"
---

# MLX Benchmarks

Structured benchmark results for **MLX-quantized** and other **locally-hosted
LLMs** on Apple Silicon. Covers throughput, time-to-first-token, tool-calling,
code generation, reasoning, knowledge, and math suites.

Results are produced by a sweep harness that wires upstream evaluation tools
against a local `vllm-mlx` inference server:

- [EleutherAI/lm-evaluation-harness](https://github.com/EleutherAI/lm-evaluation-harness) — coding, reasoning, knowledge, math
- [linusvwe/MLXBench](https://github.com/linusvwe/MLXBench) — throughput and time-to-first-token
- [vllm `benchmark_serving`](https://docs.vllm.ai/en/latest/performance/benchmarks.html) — performance second opinion
- [huggingface/lighteval](https://github.com/huggingface/lighteval) — broader task coverage

All data here is generated on Apple Silicon hardware (MINISFORUM MS-A2 / M4 Max
class), stored in flat columnar Parquet for easy querying, and appended to via
unique-filename commits so historical shards are never overwritten.

## Quickstart

```python
from datasets import load_dataset

ds = load_dataset("JacobPEvans/mlx-benchmarks")
print(ds)

# Example: average throughput per model
import pandas as pd
df = ds["train"].to_pandas()
throughput_rows = df[df.suite == "throughput"]
print(
    throughput_rows.groupby("model")["metric_value"]
    .mean()
    .sort_values(ascending=False)
)
```

Raw Parquet fetch (token-optimal for agents):

```bash
curl -sSL \
  https://huggingface.co/datasets/JacobPEvans/mlx-benchmarks/resolve/main/data/train-00000-of-00001.parquet \
  -o run.parquet
```

## Schema

Each input benchmark run produces a JSON envelope (see `schema.json` in this
repo for the authoritative v1 spec). The envelope is **exploded row-wise** into
flat scalar columns — one row per entry in the envelope's `results[]` array.
Skipped runs become a single sentinel row with null metric columns and
`skipped=true`. This mirrors the columnar layout used by the
[Open LLM Leaderboard contents dataset](https://huggingface.co/datasets/open-llm-leaderboard/contents).

| Column | Type | Notes |
| --- | --- | --- |
| `suite` | string | One of: throughput, ttft, tool-calling, code-accuracy, framework-eval, capability-comparison, coding, reasoning, knowledge, evalplus, math-hard |
| `model` | string | Full model identifier |
| `git_sha` | string | Commit SHA of the generator at run time |
| `timestamp` | string | ISO-8601 UTC start of the run |
| `trigger` | string | `schedule`, `pr`, `workflow_dispatch`, or `local` |
| `schema_version` | string | Envelope schema version (currently `"1"`) |
| `pr_number` | int64 | PR number if triggered by a pull request, else null |
| `skipped` | bool | True for sentinel rows where the suite was skipped |
| `os` | string | Operating system at run time |
| `chip` | string | CPU/chip identifier |
| `memory_gb` | int64 | Total system RAM |
| `vllm_mlx_version` | string | Backend version if captured |
| `runner` | string | Runner label or `local` |
| `metric_name` | string | Individual test/measurement name |
| `metric_metric` | string | Metric family (e.g. `throughput`, `latency`, `score`) |
| `metric_value` | float64 | Numeric value |
| `metric_unit` | string | Unit (`tok/s`, `seconds`, `ratio`, ...) |
| `tags_json` | string | JSON-serialized tag dict (per-suite custom metadata) |
| `errors_json` | string | JSON-serialized list of non-fatal errors from the run |

Nested fields from the envelope (`tags`, `errors`) are preserved as
JSON-serialized strings so no information is lost — rehydrate with
`json.loads(row["tags_json"])`.

## Update cadence

New rows are appended on every sweep via a unique-filename commit pattern
(`data/run-{timestamp}-{sha}-{suite}-{model}.parquet`). Historical shards are
never overwritten. `load_dataset()` concatenates all `data/*.parquet` files
into a single `train` split at load time.

## License

Apache 2.0 — same as the underlying upstream evaluation tools.