MCMA-20M

Validation-selected checkpoint from a 20.23M-parameter factorized Micro-Channel Attention generalist language model using the native GPT-S 4,096-token vocabulary. It has no arithmetic-specific token splitting, place embeddings, role embeddings, or inference-time equation detection. It was selected at step 1,000 for WikiText normalized BPB 1.9469. Its report-only Open SLM Leaderboard-style average is 32.90%.

Loading

This is a custom Transformers architecture. trust_remote_code=True is required because stock Hugging Face model classes do not implement MCMA or this model's exact rotary convention.

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "User01110/MCMA-20M"
tokenizer = AutoTokenizer.from_pretrained(repo, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    repo, trust_remote_code=True, dtype="auto"
)

Checkpoint tensors are stored in bfloat16. Pass dtype=torch.float32 when an FP32 runtime is required; every stored BF16 value widens exactly to FP32, though the pre-export FP32 master-weight mantissa cannot be reconstructed.

Architecture

• Parameters: 20,230,760, with tied input/output embeddings
• Weights: native bfloat16 safetensors (model.safetensors); no .bin weights
• Runtime: PyTorch 2.5+ for native SDPA grouped-query attention
• Tokenizer: AxiomicLabs/GPT-S-5M at revision df47402
• Vocabulary: 4,096 native tokens
• Parameter allocation: 1,474,560 tied embedding parameters and 18,756,200 non-embedding parameters
• Context: 1,024 tokens
• Standalone prompt tokenization automatically prepends the native BOS token
• Width/layers: 360 / 16
• Token-attention heads: 9 query, 3 KV
• MCMA: rank=64, SwiGLU hidden=720, all-channel two-GEMM mixer
• Contiguous-half RoPE without scaling
• No task-specific model or tokenizer features

Training mixture

• FineWeb-Edu 100BT shuffled: 45.00% of trained tokens
• DCLM-Baseline 1.0: 25.00% of trained tokens
• DCLM-Edu: 10.00% of trained tokens
• Cosmopedia v2: 10.00% of trained tokens
• FineMath 4+: 10.00% of trained tokens

DCLM-Baseline adds broadly curated web diversity, DCLM-Edu adds an education-filtered DCLM view, and FineMath-4+ supplies mathematical explanations and reasoning as ordinary causal-language-model text. There are no task-specific model or tokenizer features. All five training sources are streamed natural-text corpora.

Zero-shot evaluation at step 1,000

The four lm-eval tasks use normalized accuracy when supplied by lm-eval 0.4.12, with native bfloat16 weights and float32 likelihood softmax. ArithMark uses the same precision policy and its official raw continuation log-likelihood-sum rule. Autocast is not used for evaluation. Every independent benchmark context starts with the native BOS token.

Benchmark	Accuracy
HellaSwag	26.23%
ARC-Easy	29.25%
ARC-Challenge	23.81%
PIQA	52.45%
ArithMark-2	26.40%
ARC mean	26.53%
Open SLM Leaderboard-style average	32.90%

The average is (HellaSwag + mean(ARC-Easy, ARC-Challenge) + PIQA + ArithMark-2) / 4, matching the Open SLM Leaderboard formula.

WikiText-103 validation at this step: loss 4.3061, perplexity 74.15, normalized BPB 1.9469 over 358,050 scored tokens and 1,142,486 normalized UTF-8 bytes, using one initial BOS, 1,024-token windows, and a 1,024-token stride.