CoolFace
Modelpublic

Pacific-i64/TR-HASH-MOE-100M

sourceHugging Facecc-by-nc-4.0updated 2mo agoView on Hugging Face
0likes19downloads
Model Card

TR-HASH-MOE-100M

TR-HASH-MOE-100M is a 99.5M-parameter causal base language model from the Complexity TR-Hash experiments. It combines GQA with a dense shared MLP path and small residual experts selected by a deterministic balanced hash of token identity. The selected route is known without evaluating a learned contextual router.

This checkpoint is pretrained only. It is not instruction-tuned and should not be expected to follow chat prompts reliably.

Architecture

  • —Parameters: 99.5M
  • —Hidden size: 384
  • —Layers: 10
  • —Attention: GQA, 8 query heads / 2 key-value heads
  • —Vocabulary: o200k, 200,019 entries
  • —Context length: 2,048
  • —Shared MLP width: 1,520
  • —Residual experts: 4 experts, top-2, 0.5 / 0.5 route weights
  • —Routing: deterministic balanced token-ID hash
  • —Hash channel modulation: enabled
  • —Tied token embeddings / LM head
  • —Weight dtype in this release: BF16

Training evidence

The model was trained for approximately 2B FineWeb-Edu tokens with seed 42. Training used a separate fixed evaluation shard. At the 10,000-step evaluation, TR-Hash reached 3.6680 NLL; the matched Dense-GQA reference reached 3.6685. Measured training throughput was about 375.8k tokens/s for TR-Hash and 397.5k tokens/s for Dense-GQA on the recorded run.

These are single-seed experimental measurements, not evidence of a general quality or speed advantage.

Runtime

The exported checkpoint uses the DeepForCausalLM contract implemented in the Complexity vLLM fork. Standard Transformers does not natively implement this TR-Hash architecture.

  • —Training and architecture: https://github.com/Complexity-ML/complexity-framework
  • —CUDA inference: https://github.com/Complexity-ML/vllm-cuda_graph

The included tokenizer metadata points to the o200k_base encoding. This base checkpoint deliberately contains no chat template.

Intended use and limitations

This release is intended for architecture research, evaluation, continued pretraining, and controlled fine-tuning. A 99.5M-parameter base model has limited factual knowledge and generation quality. Outputs may be incorrect, repetitive, biased, or unsafe and should not be relied on for consequential decisions.

License

Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0).