以下部分引用 AI 总结,现阶段 AI 仍然有幻觉。请以内容的原文为准。
数据新鲜度群聊 07-14 ✓卖方 07-14 ✓Wrap 断档15天 ✕News 07-15 ✓X 断档12天 ✕页面生成 07-15
poolside
8 atoms · 跨 4 天 · 首见 2026-04-28 · 最近 2026-07-02
三色: 🟦 fact 7 · 🟥 take 1 · stance ▲0/▼1/◆2
来源: X 8
时态: fresh:7 · aging:1
标签: 好思考:4 · 好信源:3 · 好问题:1 · 好数字:1
叙事 (narrative) (1)
- 2026-04-28
[老化中] · 使用了 moe_router_logit_softcapping 参数
_there is a "moe_router_logit_softcapping" parameter in the config set to 0 by default, but this is non-standard i think, means that they likely saw some instabilities in the router scores?_
tags: 好思考·好问题 · → daily
value: qty=0 · date=default
事实 (fact) (7)
- 2026-07-02 · Laguna XS 2.1 发布
Today we’re releasing Laguna XS 2.1. It’s a small upgrade to the Laguna XS.2 model, the same 33B total / 3B active MoE and stronger results on multilingual coding and terminal-style tasks.
tags: 好信源 · → daily
value: qty=33B total / 3B active MoE · date=2026-07-02
- 2026-06-29 · forked from torchtitan
ye and poolside also forked from torchtitan i believe
tags: 好信源 · → daily
- 2026-06-29 · AI实验量
100 experiments a week
tags: 好数字 · → daily
value: qty=100 · date=每周
- 2026-05-26 · 发布Laguna模型技术报告
Today we’re publishing the technical report behind Laguna M.1 and Laguna XS.2.
tags: 好信源 · → daily
value: date=2026-05-26
- 2026-04-28 · 使用了 per head gated attention
they use per head gated attention like qwen or stepfun, people usually use sigmoid, but they use softplus! they also use it on both swa and full attention which is nice!
tags: 好思考 · → daily
- 2026-04-28 · SWA 层使用了更多 attention heads
more attention heads on swa layers, xiaomi v2 flash did higher kv heads for swa layers
tags: 好思考 · → daily
- 2026-04-28 · YARN 的超参 beta_fast 设为 64
_they use a different beta_fast for yarn than most models (64 instead of 32) and partial (half) rope for the full attention layers_
tags: 好思考 · → daily
value: qty=64
时间轴 (近 20)
- 2026-07-02 ·
fact · Laguna XS 2.1 发布 · → daily
- 2026-06-29 ·
fact · forked from torchtitan · → daily
- 2026-06-29 ·
fact · AI实验量 · → daily
- 2026-05-26 ·
fact · 发布Laguna模型技术报告 · → daily
- 2026-04-28 ·
fact · 使用了 per head gated attention · → daily
- 2026-04-28 ·
narrative · 使用了 moe_router_logit_softcapping 参数 · → daily
- 2026-04-28 ·
fact · SWA 层使用了更多 attention heads · → daily
- 2026-04-28 ·
fact · YARN 的超参 beta_fast 设为 64 · → daily
← 实体目录 · 系统日志