AmigaMeow/llm-leaderboard

LLM Leaderboard (Arena x API pricing)

Daily-updated LLM leaderboard combining LMArena human-preference (Arena) scores with OpenRouter public pricing, so you can answer "what should I use at $X?" — a question neither source answers on its own.

Source of truth: 17nas.com/llm-leaderboard.php ｜ Upstream repo: AmigaMeow/llm-leaderboard-data

Files

Path	内容
data/models.jsonl	推荐入口：一行一个模型（扁平结构，可直接 load_dataset）
data/latest.json	最近一次快照的原始结构（含 sources 元信息）
data/history/YYYY-MM-DD.json	按天归档的历史快照，可做差分

Fields

• id / display_name / org — 模型标识
• arena_score / arena_ci_low / arena_ci_high / arena_votes / arena_rank — LMArena Bradley-Terry 评分与 95% 置信区间。区间重叠时名次差没有统计意义， 请据此判断，不要只看名次。
• price_in / price_out — OpenRouter 报价，美元 / 百万 token
• blended_price — 便捷派生值 (3 * price_in + price_out) / 4（假设输入:输出 = 3:1， 按你自己负载调整）
• context_length / open_weights / license / hf_repo / listed_at_iso / sources_hit
• snapshot_date / snapshot_generated_at — 本行所属快照
• lmarena_publish_date — 上游 Arena 榜的发布日期，与抓取日期是两回事，见下

Important: two different dates

snapshot_date is when this repository fetched the data. lmarena_publish_date is when LMArena published the leaderboard snapshot being used.

LMArena publishes on its own schedule — there can be long gaps. Always check lmarena_publish_date before treating Arena scores as current.

Usage

from datasets import load_dataset
ds = load_dataset("<this-dataset>", "models", split="train")
print(ds[0])

Attribution and licensing

• LMArena leaderboard data — licensed CC-BY-4.0; attribution required (given here).
• OpenRouter public pricing API.
• Artificial Analysis is deliberately NOT included — their terms prohibit redistribution.
• This compilation is published under CC-BY-4.0 to match the most restrictive upstream source. See the upstream repo's docs/UPSTREAM-TOS.md.

Update cadence

Rebuilt daily by a GitHub Actions workflow right after the upstream repository fetches fresh data. See data/history/ for per-day snapshots.