#Mistral-7B

2 stories · 1 in 7 days · 1 in 30 days

KV缓存 1 Llama-3.1-8B 1 Llama-3-8B 1 Prometheus 1 Qwen2.5-7B 1 Themis 1 推理优化 1 评分模型 1

Stories, products and papers tagged "Mistral-7B", newest first.

Oct 27

Papers Builders10:25

Chinese contentBellowsKV：跨层共享 KV 缓存，压缩 13 倍仍保留 97% 精度

免训练的 KV 缓存压缩方案，压 13 倍几乎不掉点，比现有跨层方法强不少，自己部署 LLM 的可以看看代码。

#KV缓存 #Llama-3.1-8B #Qwen2.5-7B #Mistral-7B

arXiv cs.LG@Jiamu Zhang, Hanjie ChenSource

Sep 2

PapersSelected73°10:15

LLM-as-a-Judge Mechanisms in Summarization

Fine-tuning sculpts existing substrates rather than building pipelines from scratch, with suppression of below-L15 MLP contribution and two-layer advance of crystallization depth.

#Themis #Prometheus #Llama-3-8B #Mistral-7B

arXiv cs.LG@Himil Vasava, Ming JiangSource