LEM-Eval

The 8-PAC benchmark runner for the Lemma model family.

A HuggingFace dataset repo used as a tool-shaped "github" — the entire scorer lives here, anyone clones it, installs once, and the worker machines chug along advancing per-model canons in lockstep. Multiple workers farm different targets in parallel — each target declares a type (mlx or gguf) in targets.yaml, and workers filter by the backends they can actually run (capability probe or LEM_TYPES env). Partition falls out of what the hardware can do, not hostnames.

What it does

For each declared target (a base + LEK-merged model pair), runs a paired 8-PAC benchmark:

• 8 independent rounds per question using Google-calibrated Gemma 4 sampling (temp=1.0, top_p=0.95, top_k=64, enable_thinking=True)
• Both models see the exact same question set from the seed-42 shuffled test split — the only variable is the weights
• Auto-offset progression: each run advances the canon by n_questions via lighteval's --samples-start flag, so consecutive runs naturally cover contiguous non-overlapping windows

Results are written to two canonical destinations per run:

1. The target model repo's .eval_results/<task>.parquet — primary, per-model scorecard, drives the HF model-card eval_results rendering
2. lthn/LEM-benchmarks/results/<target>/<task>.parquet — aggregated, fleet-wide, grows as more machines contribute observations

Same row data, two locations. Dedup on (machine, iter_timestamp, question_index, round, model_side) on both canons, so the same machine re-running the same slice is idempotent but different machines contribute additive rows to the aggregator.

Layout

LEM-Eval/
├── eval.py               # target-driven runner (PEP 723 — uv run it)
├── mlx_lm_wrapper.py     # lighteval custom model backend
├── targets.yaml          # declarative fleet spec (base, this, type)
├── install.sh            # bootstrap: clone model repos + lem-benchmarks
├── lem-eval.sh           # service script (once | maintain | loop)
├── cron/
│   ├── submit.cron       # */30 * * * * lem-eval.sh once
│   └── maintain.cron     # 15 * * * * lem-eval.sh maintain
└── workspaces/           # local clones of target model repos (gitignored)
    ├── lemer/
    ├── lemma/
    └── ...

Quick start (worker)

export HF_TOKEN=hf_...    # or: huggingface-cli login
git clone https://huggingface.co/datasets/lthn/LEM-Eval
cd LEM-Eval
./install.sh              # clones lem-benchmarks + your owned model repos
./lem-eval.sh once        # run one pass manually to verify

# Install the continuous cron
crontab -l | cat - cron/submit.cron cron/maintain.cron | crontab -

Add a new machine: install LEM-Eval on it, the worker's backend probe decides which targets it can run (mlx on Apple Silicon, gguf where an Ollama endpoint is reachable). Override with LEM_TYPES=mlx,gguf in the cron env if you want explicit control. Workers pick up targets.yaml edits via the maintain cron's hourly git pull.

gguf wrapper status: not yet implemented. gguf targets (lemmy, lemrd) sit in targets.yaml waiting for gguf_wrapper.py — will be an OpenAI-SDK wrapper pointing at a local Ollama/llama.cpp server. Until then, gguf targets list but don't run.

Quick start (manual / dev)

uv run eval.py --list-targets                     # show the fleet
uv run eval.py --my-targets                       # show targets owned by $(hostname)
uv run eval.py --target lemer --n-questions 1 --rounds 8
uv run eval.py --target lemer --loop 8            # 8 back-to-back advances

PEP 723 inline metadata declares all dependencies — no venv setup needed. uv creates one automatically, caches it, and pulls lighteval from our fork which carries benchmark-stability patches (MMLU-Pro template fix, --samples-start offset).

Related

• lthn/LEM-benchmarks — aggregated results store
• LetheanNetwork/lighteval — benchmark-stability fork
• lthn/lemer, lemma, lemmy, lemrd — target models

License

EUPL-1.2.