Unverified50% confidenceFactExact time
The author's publicly shared standard training configuration is: Krea 2 Standard model, Standard LoRA, Network rank 32, Alpha 32, Epochs 64 (saving every 4 epochs), Batch size 1, target resolution 0.25 MP, EMA 0.98.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
9/24/2026
First Seen
Valid until: 12/23/2026
Sources
Related Entities
Related Claims
UnverifiedLoRA训练中常用的学习率为1e-4到5e-4之间，常用的Network Rank值为8、16、32、64、12866% similarUnverifiedLoRA使可训练参数量通常仅为全参微调的0.1%-1%，可在单张RTX 3090/4090（显存24GB）上微调7B乃至13B级别的模型62% similarUnverifiedKimi K2.7 Code uses a MoE architecture with 384 experts, selecting 8 experts plus 1 shared expert per token, with actual computation approximately 2.3% of that of a dense model.61% similarUnverifiedRank-256 is higher than the commonly used rank-8/16, meaning more trainable parameters and a greater degree of stylistic modification to the original model.60% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/948325API
curl https://kongchang.com/api/v1/knowledge/claims/948325MCP
get_claim(id=948325)