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

Talkie

9 atoms · 跨 2 天 · 首见 2026-06-01 · 最近 2026-07-02

三色: 🟦 fact 7 · 🟥 take 2 · stance ▲1/▼1/◆0

来源: X 9

时态: fresh:7 · stale:2

标签: 好数字:3 · 好观点:3 · 好信源:3

展开 2 条老旧 / 已过期

2026-06-01 [老旧] · 32k上下文扩展性能
It's probably possible to squeeze some more long context performance from Talkie, but for two afternoons on a single A100 node I think this is reasonable, given where the model started
tags: 好观点 · → daily
value: qty=reasonable · date=2026-06-01
2026-06-01 [老旧] · YaRN扩展方案建议
The Talkie folks told me it actually was 4k, so I tried again with 8x from 4k seqlen, but the performance was worse so I recommend the 16x extension
tags: 好观点 · → daily
value: qty=16x · direction=推荐

2026-07-02 · 训练数据截止年份
Talkie, a model trained only on text from before 1931
tags: 好信源 · → daily
value: qty=1931 · date=before 1931
2026-07-02 · 对上帝的看法
thinks God is extremely important
tags: 好信源 · → daily
value: direction=extremely important
2026-07-02 · 国籍认同
is “very proud to be a citizen of Great Britain”
tags: 好信源 · → daily
value: direction=very proud to be a citizen of Great Britain
2026-06-01 · context长度扩展
I ran a quick 32k context extension on Talkie 1930 using pre-1931 books and YaRN
tags: 好数字 · → daily
value: date=2026-06-01 · qty=32k
2026-06-01 · 原始seqlen
The Talkie reference code showed a 2k seqlen
tags: 好数字 · → daily
value: qty=2k
2026-06-01 · checkpoint性能对比
the 2k 16x step500 checkpoint meaningfully outperformed the final step1000 checkpoint with a fully decayed LR
tags: 好数字 · → daily
value: qty=2k 16x step500 · direction=优于step1000
2026-06-01 · 模型类型
Remember that this is a base model, so will need fine-tuning to function as a full chat model
tags: 好观点 · → daily
value: qty=base model