TR-HASH-MOE-500M-HF — Card Corpus V2 epoch 1

Experimental merged checkpoint from step 963, the end of the first Card Corpus V2 LoRA epoch. The LoRA weights are already merged; no separate adapter is required.

Training state

Component	Value
Base	AETHORIA-AI/TR-HASH-MOE-500M-20B
Dataset	Complexity Card Corpus 32K V2 trial build
Training examples	184,801
Checkpoint	963 / 963 steps (1 epoch)
Final training loss	1.2775
LoRA	rank 24, alpha 24, dropout 0.05
Peak base LR	1e-5
Peak expert LR	2.5e-6
Chat format	complexity-chat-v2

The bundle contains merged BF16 weights, the 32K tokenizer, the standalone chat template, and the fixed layer-specific top-2 route tables.

Inference

pip install git+https://github.com/Complexity-ML/vllm-i64.git@main

vllm-i64 serve tr-hash-moe-500m \
  --checkpoint AETHORIA-AI/TR-HASH-MOE-500M-HF \
  --dtype bfloat16 \
  --port 7860

Limitations

Token base pretraining limitation. This is a diagnostic research checkpoint, not a production assistant. The first epoch improves assistant-like responses over the pretrained base, but controlled smoke tests still show unreliable arithmetic, instruction following, safety responses, and synthetic response patterns. It has not passed the promotion gate and should be independently evaluated before use.

Resources

• Training framework
• Custom runtime
• Post-training dataset