Spaces:

---

AxiomicLabs

/

Open_SLM_Leaderboard

Like 145

Running

Model submission: MICA Flame-W 0.3.6

Proposes one model row and one maker entry. Scores are self-reported pending your verification.

Model: https://huggingface.co/vynly/mica-flame-w-0.3.6
Published checkpoint SHA256: eca9faee6f44e3b2e4c10399ec99d202fa941a10ef78e64fb2094b0de70a4616.
Pinned reproduction code, per-item score evidence and data/model hashes.

Architecture, eligibility and comparability disclosures

MICA is a trained-from-scratch integer-rule cellular automaton with a NumPy CPU runtime, not a Transformer. This entry uses its word language model, integer short/long memories, frozen answer-mode calibration, previous-word bank and phrase-copy table. No English belief-memory adapter or choice override is used. The exact configurations are in PROTOCOL.md.

The automaton uses training-time averaging of fitted rule immediates/readouts within one live training trajectory, not a merge of independently pretrained models. All artifacts are publicly downloadable under the MICA AI Non-Commercial License (research/personal use; not OSI open source).

Because this is non-neural, the proposed params field is explicitly a conservative stored-value proxy: 31,523,432 numeric scalars including fixed wiring and metadata, displayed as “31.52M values.” It is not a claim of neural trainable-parameter equivalence. Please review whether this fits your sizing/checkpoint rules before merging; happy to use a separate non-neural category if preferred. No inflated rank or architectural superiority is claimed.

Protocol and reproduction

Zero-shot, canonical lm-eval 0.4.13 task prompts/preprocessing/acc_norm for the first four tasks; local JSON loaders contain the exact pinned official full splits. ArithMark-3 follows its linked script's model-token normalization (MICA tokens are words). Tokens casefold text and hash OOV words into 256 buckets; this is native word-symbol scoring, not lossless byte compression.

The harness reproduction uses the frozen, independently reproduced release's native per-choice likelihood cache, keyed by complete request hash and bound to every package file hash. Greedy flags are computed live; 67 likelihoods were systematically recomputed live, plus an all-live smoke check. The full-split harness counts reproduce all five published results. This is disclosed cache reuse, not a second complete uncached evaluation. The cache is optional; for independent verification run without it:

pip install lm_eval==0.4.13 numpy pyarrow torch
# obtain the pinned Flame-W package and clone the linked reproduction directory
python fetch_slm_data.py
python flame_lm_eval.py --package /path/to/flame-package --data data --workers 8

Torch is only a harness dependency; MICA inference remains NumPy and integer cellular rules. No training or score selection was performed for this submission. These are public benchmarks previously inspected during research; we do not claim a never-inspected blind evaluation. Standard errors and per-item evidence are in the linked artifacts. There is no ArithMark-2 score; no missing result is filled with zero.

I checked maintenance before submitting: your Space merged model entries on October 12, 11, 10, 9 and 8. Thank you for maintaining it and for reviewing this native-runtime entry.