quaedra/local-llm-evals

Local LLM evals

The evidence behind Best Local LLM for Your GPU: every benchmark score the index considered, where it came from and how much it is trusted, plus raw logs of our own evaluation runs.

scores.jsonl

One row per score, rebuilt with the index.

field	meaning
model	the Hugging Face repo the index ranks
benchmark	arena, arena_coding, gpqa, mmlu_pro, hle, aime_2026, swe_verified, swe_pro, terminal_bench, aa
value	the score as published
provenance	independent (measured by a third party), harness (our runs), self (model card or paper)
source, source_url	where it was published
reported_for	the repo the score was reported for (a quantized or renamed repo of the same model is grouped with it)
estimated	true when Artificial Analysis marks the score as its own estimate
built	when the index was built

How the index weighs these: independent 1.0, our harness 0.85, self-reported 0.5, times how well each benchmark predicts LMArena (r²). An index resting on fewer than two independent measurements is marked provisional.

runs/

Raw per-question results of our own runs (one JSONL per model build and benchmark, plus the command, hardware and llama.cpp version), starting with MMLU-Pro retention of quantized builds against their full-precision reference (retention_eval.py in the Space).

Sources

LMArena (lmarena-ai/leaderboard-dataset), Hugging Face benchmark leaderboards (EvalEval, official MathArena / SWE-bench / Terminal-Bench evaluations, and model cards), and the Artificial Analysis Intelligence Index as recorded by the Sybil local-ai-registry. Scores remain their publishers'; check the source before relying on one.