GoLLeM-v5 64M Muon v1 (earlier recipe)

First 64M flagship of the series (Qwen3-style decoder, Muon optimizer). Earlier recipe than the final models (13.1B instead of 24.9B tokens); the final 64M model supersedes it on eff by +0.26.

Author: Arkadiusz Słota / Fabryka AI. Part of the GoLLeM-v5 series of small English language models trained from scratch. All checkpoints of the series, including intermediate ones, are archived in SlayerLab/gollem-v5-ckpts.

Model

• 62.9M parameters; 14 layers, d_model 576, 9 heads (head dim 64); context 1024 tokens.
• Architecture: Qwen3-style decoder: RoPE (theta 100,000), SwiGLU (FFN multiplier 2.667), RMSNorm, QK-norm, value residual; tied input/output embeddings.
• Tokenizer: BPE, 12,288 tokens (tokenizer.json, sha256 3733307577230bb4802d2c774d8e2323f7f64e4139c13712736a57daf91bdda1), 3.8605 bytes per token on WikiText-2 test.
• Weights: model.safetensors (its sha256 is shown by the Hub for the LFS file), converted tensor by tensor from v1_muon/ckpt_400k.pt (sha256 59f982c1516c6e19df86d4ee1ac6359cd1987830a624080c70d07db0c2d0455f) without changing any weight. Only weights are published here: no pickled checkpoint, optimizer state or training logs.

Training

• Recipe: Muon (hidden 2-D weights, lr 0.02) + AdamW for non-matrix parameters, cosine schedule 6e-4 to 6e-5, batch 32 x 1024 tokens, 400,000 steps = 13.1B tokens (~1.4 epochs over ARC-MIX), seed 1337.
• Data: ARC-MIX (9.42B BPE tokens): the base mixture SlayerLab/minimal-en-corpus-5b (~5.4B tokens: FineWeb-Edu, DCLM, StackExchange, open-web-math, FineMath, scientific papers, books/Gutenberg, code, CC-News) plus FineWeb-Edu and OpenStax textbooks, with ARC-relevant science, reasoning and Q&A web content upweighted (gold ~3x, related ~2x). Filtered against benchmark test sets with 13-gram matching at build time.
• Result checkpoint: The result is the last step of the planned run (400,000).
• Training run (loss curves, metrics): https://track.fabryka.ai/run/16b892d3-2a72-4b54-a595-6f03a9ad4a8b
• An independent re-benchmark by the board maintainer on the same checkpoint (sha256 matched) reproduced ARC-Easy 47.94 exactly.
• In a clean 64M A/B (same data, architecture and seed) Muon was at least as good as AdamW.

Evaluation (our measurement, not official)

Measured by us on the result checkpoint with the Glint Tiny-ML board protocol (BLiMP: 67 configs, 67,000 pairs, sentences clipped to 256 tokens, raw log-prob preference; ARC-Easy: test split, zero-shot, LL(question + choice) - LL(question); WikiText-2: byte-normalized perplexity). These are not official scores.

• eff = mean of BLiMP, ARC-Easy and normalized WikiText-2, times a size multiplier that is larger for smaller models; compare eff only between models under the same board formula.
• This is an English model. Its Polish MultiBLiMP score is close to the length baseline (always choosing the shorter sentence gives 59.67 %); it is reported for completeness, not as Polish ability.
• Board: https://track.fabryka.ai/models (row "GoLLeM-v5-64M-Muon-v1").
• Source of the numbers: liczby_v5.json (sha256 35c9022ee9dbd96d588093c6fc69604d0eaabd18843458671205db41e234dde8), built from the anonymous board read-out and our PL measurement files.

Benchmark overlap disclosure

A scan of the finished training corpus found that 7 of 62 WikiText-2 test articles appear almost completely in it (as web copies) and 30 of 2,376 ARC-Easy test questions have a match (mostly short factual sentences). On the 16M, 32M and 64M Muon v1 models, removing them changed WikiText-2 byte perplexity by less than 0.01 and ARC-Easy by -0.07 to +0.19 points, within noise (details in the root card of the archive repository). BLiMP was covered only by the build-time 13-gram filter.

How to use

The weights are a plain PyTorch state dict saved as safetensors; they are not a transformers AutoModel. The model class is GPT in train_gpt_ref.py in SlayerLab/gollem-v5-ckpts; build it from config.json and load the state dict, then trim logits to 12,288.

Limitations

• Small base model trained from scratch for research: not instruction-tuned, limited factual knowledge and coherence, not for production use.
• English only.
• Single seed; scores come from one checkpoint and our own implementation of the board protocol.

Training data

Base mixture: SlayerLab/minimal-en-corpus-5b (research-mix-5b, 5.0B tokens of its own tokenizer). It is an aggregate corpus: no unified licence is applied and each source keeps its upstream terms. Sources as named in its manifests/mixture.json:

• Upstream repositories are not recorded in the corpus manifest; licences are those of the Hub datasets matching these source names (not confirmed against the build).
• Added on top of the base mixture for this series: FineWeb-Edu (ODC-By 1.0) and OpenStax textbooks (CC BY 4.0).
• Aggregate corpus; each source keeps its upstream terms. Commercial use: review upstream terms, in particular CC-News, scientific papers, StarCoderData and model-generated chat data.

Licence and attribution

• Weights: no single licence is claimed for the weights yet; they were trained on data under the licences listed in the Training data section above, and any use must respect those terms (license id other, name gollem-v5-see-training-data).
• OpenStax textbooks: CC BY 4.0, © Rice University, https://openstax.org; titles and editions in OPENSTAX_ATTRIBUTION.md in this repository.