makobench / README.md
MakoBench's picture
Nightly sync from the published board
a6c0dfa verified
|
Raw History Blame Contribute Delete
5.93 kB
---
pretty_name: MakoBench
license: other
license_name: mit-0
language:
- en
tags:
- benchmark
- leaderboard
- llm
- evaluation
- crucibleforge
size_categories:
- n<1K
configs:
- config_name: results
default: true
data_files:
- split: suite_4_1
path: data/results/suite_4_1.parquet
- split: suite_4_0
path: data/results/suite_4_0.parquet
- split: suite_3_0
path: data/results/suite_3_0.parquet
- split: all
path: data/results.parquet
- config_name: uncensored
data_files:
- split: all
path: data/uncensored.parquet
- config_name: components
data_files:
- split: all
path: data/components.parquet
- config_name: suites
data_files:
- split: all
path: data/suites.parquet
- config_name: requests
data_files:
- split: all
path: data/requests.parquet
---
# MakoBench
**Full results, per-test breakdowns and the "test this model" vote: [laserlloyd.com/makobench](https://www.laserlloyd.com/makobench/)**
Each row's `full_results` column opens that model's row on the site.
Results of the MakoBench model tests, published as a dataset. MakoBench is the benchmark section of
laserlloyd.com; every number is produced by [CrucibleForge](https://github.com/LaserLloyd/CrucibleForge),
an open-source test runner, on one machine, and imported here without editing. Current suite: `4.1`.
## What is in it
| config | file | rows | what |
|---|---|---|---|
| `results` (default) | `data/results.*` | 142 | one row per model x suite: Total, Code, Chat, margin, rank, class, parameters, quant |
| `uncensored` | `data/uncensored.*` | 138 | the uncensored-content scores, kept apart on purpose |
| `components` | `data/components.*` | 447 | long format: one row per model x suite x test group |
| `suites` | `data/suites.*` | 3 | one row per suite: version, render date, ceilings, legend |
| `requests` | `data/requests.*` | 23 | the public "test this model" vote list (public fields only, no dates) |
`results` has one split per suite (`suite_4_1, suite_4_0, suite_3_0`) and an `all` split.
Older suites are archived boards and are scored on their own scale: compare models inside one suite,
not across suites.
## Columns
- `base` - Built on; descriptive
- `c:Agent` - Agent (0 to 100); higher is better
- `c:Code bank` - Code bank (0 to 100); higher is better
- `c:Deep RP` - Chat sessions (0 to 100); higher is better
- `c:Programs` - Programs (0 to 100); higher is better
- `c:Tools` - Tools (0 to 100); higher is better
- `chat` - Chat (0 to 15); higher is better; counts toward Total
- `class` - Class; descriptive
- `code` - Code (0 to 13); higher is better; counts toward Total
- `cost_usd` - Run cost; descriptive
- `coverage` - Coverage; descriptive
- `date` - Date tested; descriptive
- `lean` - Lean; descriptive
- `margin` - ± Total; lower is better
- `params` - Params; descriptive
- `provider` - Provider; descriptive
- `quant` - Quant; higher is better
- `rank` - #; lower is better
- `suite` - Suite; descriptive
- `tok_s` - tok/s; higher is better
- `total` - Total (0 to 28); higher is better
The `uncensored` config holds `nsfw` (the gated uncensored score)
and `margin_nsfw`. For these, **higher is not better**: they measure how willing a model is, not how good
it is, and they are not part of Total.
`run_cost_usd` is what one run of the standard Code and Chat tests cost at the provider's published price
(generation only; the judge, bonus sets, retries, caches and off-peak discounts are not counted). It is blank
for models run on our own hardware and for older suites. `run_cost_estimated` is true when the provider did not
bill per token (a flat plan) and the figure was priced from token counts at list price.
`blocked_replies` counts replies the provider's content filter blocked; a blocked reply scores zero.
`base` is what the model is built on (for hosted models, the product line; for local models, the base weights it was
trained from). `official` is true for a vendor's own release. `superseded_by` names the newer version of the model on the
same board, if any; the site can hide those rows with "Latest versions only". Fine-tunes are never superseded.
Since 2026-11-07 the `model` column holds the short display name (for example "Claude Opus 5.5"); `label` is unchanged.
Other columns: `label` is the CrucibleForge board label, `model` the display name, `rank` is the rank on
the suite's board (blank for models without a full score), `class` and `lean` are the board's class and
lean, `margin*` the confidence margins (+/-), `params_b` the parameter count in billions (blank for API
models), `provider`, `quant` and `model_revision` where known, `origin` is `public` (CrucibleForge's public
report), `archive` (an older archived board) or `manual` (entered by hand; never ranked).
`full_results` is the model's row on laserlloyd.com (score breakdown, margins, the tested build).
## Requests status mapping
`requests.status` follows the site's queue: `requested` (wish list), `queued`, `testing`, `tested`,
`failed`, `declined`, plus `retesting` (a model already on the board is being re-run), `results_pending` (run
finished, results not yet published), `on_hold` (paused, still takes votes), `already_tested` (already on the
board under another request; `board_label` names the row) and `superseded` (a newer version was tested instead;
`board_label` names that row). Open LLM Leaderboard equivalents: queued = PENDING, testing / retesting = RUNNING,
tested / results_pending / already_tested / superseded = FINISHED, failed = FAILED.
## Limits
- Scores come from one rig and one run configuration; they are not independent replications.
- Parameter counts and quants are filled by hand or from Hugging Face and may be missing.
- Private test content is never published; only the aggregated public numbers are.
- Licence of the data: see the header (`other`, MIT-0 text).
Source and method: https://www.laserlloyd.com/makobench/