---
pretty_name: Hemmingway-1 oMLX Quantization Benchmark
language:
- en
task_categories:
- text-generation
tags:
- benchmark
- llm-evaluation
- quantization
- mlx
- apple-silicon
- model-comparison
size_categories:
- n<1K
---

# Hemmingway-1 oMLX Quantization Benchmark

This is the public-safe benchmark package for the Hemmingway-1 oMLX
quantization study on Apple Silicon.

[Altworld](https://huggingface.co/Altworld) developed and published
[Hemmingway-1](https://huggingface.co/Altworld/Hemmingway-1). Bobby Pierce
published these quantizations and the evaluation package. The
[collection](https://huggingface.co/collections/sixstringzen/hemmingway-1-omlx-oqe-quantizations-apple-silicon-evidence-6ab06103552bc3c7e9ad8afa)
links the upstream model and all six builds.

Analysis revision 2, corrected on 2026-09-22, fixes A/B attribution and matching
across reversed packets. Read [CORRECTION.md](CORRECTION.md) before using the
aggregate results. The study has 15 prompts per quant and 11 hosted comparison
prompts. All generations and judge records are unchanged.

The release contains the authored task prompts, selected local execution
metadata, aggregate blind-judge results, reliability metadata, and the policy
used to select records when a condition was run more than once.

## What is in the dataset

| File | Rows | Purpose |
| --- | ---: | --- |
| `data/train.jsonl` | 184 | Mixed rows. Filter `record_type` for task, local execution, hosted reference, or aggregate judge result |
| `metadata/reliability.yaml` | 1 object | Agreement, order-swap, and position-bias metadata |
| `metadata/selection_policy.yaml` | 1 object | Duplicate-run and future fidelity-study rules |
| `metadata/gemini-3.8-flash-veracity-review.yaml` | 1 object | Historical package review, superseded for aggregate accuracy |
| `metadata/benchmark_manifest.yaml` | 1 object | Schema, counts, provenance, and publication boundary |

The Dataset Viewer reads `data/train.jsonl` as the default configuration. The
YAML files are metadata and are not benchmark rows.

The execution records contain response hashes, lengths, status, token counts,
latency, time to first token, throughput, sampler settings, and model
condition. Response text is deliberately not included.

Use `record_type` to select the row family. Task rows contain the prompt text.
Local execution rows contain selected run metadata. Hosted reference rows are
quality-only. Aggregate judge rows contain the pooled quality results.

The `interval_percent` column is retained for schema compatibility and is null
in revision 2. The previous intervals treated dependent judgments as independent
samples and have been withdrawn. Explicit counts and judge-level totals are in
`metadata/public_results.yaml`. Task `quality_max_tokens` records the final cap
selected for judging; execution `max_tokens` records each attempt's actual cap.

## Record selection

The clean BF16 512-token rerun is the canonical reference. The earlier BF16
run contains the same 11 logical examples, response hashes, and record IDs, but
its runtime summary differs. It is excluded from merged quality and runtime
data. Runtime measurements from repeated attempts are never averaged.

The 512, 1024, and 2048 token caps remain separate strata. A logical key is
based on source run, condition, prompt, cap, seed, repetition, chat-template
settings, and runtime profile. `record_id` is retained as audit metadata, not
as the primary deduplication key.

## Provenance

The local measurements were produced by real MLX/oMLX software on an Apple M5
Max machine with 128 GB of memory. Claude monitored or orchestrated a subset
of the workflow. Claude is not the inference engine, benchmark implementation,
or measurement source.

Grafana is a derived visualization layer. The structured local manifests,
generation records, summaries, state files, attestations, and telemetry are the
source evidence for the study.

## What is not included

This public release does not contain raw local model outputs, raw provider
responses, judge packets, telemetry, request bodies, private answer keys, or
unredacted benchmark generations. Those remain in the local evidence package.

KLD, top-k agreement, KV-cache divergence, broader task scores, and controlled
runtime comparisons are pending. They are not inferred from this release.

## Reproduction

Load `data/train.jsonl`, filter `record_type` to `task` and
`local_execution`, preserve the prompt text and task IDs, and run each model
condition with the settings recorded in `metadata/benchmark_manifest.yaml` and
`metadata/selection_policy.yaml`. Keep each token-cap stratum separate.

This release publishes the task prompts, so it is a reproducibility package,
not a contamination-resistant hidden evaluation set. Use a private holdout for
future claims about generalization.

## Validation

The package is checked for row counts, logical-key uniqueness, excluded fields,
and selection-policy invariants. Analyzer regression tests exercise A/B decoding,
reversed prompt order, result validation, and packet integrity. The earlier
Gemini package review did not detect the analysis errors. Its historical PASS
does not validate revision 2, and no new model-judge review was run for this correction.

## License

No reuse license is specified in v1. The task prompts are published for
benchmark reproducibility. Reuse terms will be added in a later version.