Embeddings Are Overpriced? A Frozen 16 MB Vocabulary for Small and Ternary Language Models
Computer Science > Computation and Language arXiv:2610.22428 (cs) [Submitted on 15 Oct 2026] Title: Embeddings Are Overpriced? A Frozen 16 MB Vocabulary for Small and Ternary Language Models Authors: Sai Kiran Bathula.
计算机科学 > 计算与语言 arXiv:2610.22428 (cs) [提交于 2026 年 10 月 15 日] 标题:嵌入层定价过高?用于小型和三元语言模型的 16 MB 固定词汇表 作者:Sai Kiran Bathula。
Abstract: In a small language model the vocabulary is the single largest part. In Gemma 3 270M the embedding table holds 168M of 268M parameters, and because quantisation methods leave the embedding in 16-bit, it is also most of the file once the rest of the model is compressed.
摘要:在小型语言模型中,词汇表是最大的单一组成部分。在 Gemma 3 270M 模型中,嵌入表占据了 268M 参数中的 168M,并且由于量化方法通常将嵌入层保留为 16 位,因此在模型其余部分被压缩后,它也占据了文件的大部分空间。
We ask what that table actually buys. We replace it with one frozen table of 512-bit codes, 16 MB, with no trainable parameters, derived once from a pretrained embedding by principal-component projection and sign binarisation, and we train the same Gemma-shaped body from scratch on the same 2.05B tokens with and without it.
我们探讨了该表实际带来的价值。我们用一个 16 MB 的 512 位代码固定表替换了它,该表没有可训练参数,是通过主成分投影和符号二值化从预训练嵌入中一次性导出的。我们使用相同架构的 Gemma 模型,在相同的 2.05B 个标记上,分别在有无该表的情况下从头开始训练。
The learned table buys 0.20 nats of validation loss for 2.6 times the trainable parameters, but on LAMBADA, HellaSwag and PIQA the frozen-table model is better or tied; only ARC-Easy favours the learned table.
学习型词汇表以 2.6 倍的可训练参数换取了 0.20 nats 的验证损失提升,但在 LAMBADA、HellaSwag 和 PIQA 测试中,固定表模型表现更好或持平;只有 ARC-Easy 测试更倾向于学习型词汇表。
With a ternary BitNet b1.58 body the whole model is a 46 MB file, still ties the float model on LAMBADA and PIQA, and on a laptop CPU generates 615 to 740 tokens per second in 62 MB of memory, five to six times the speed of the float model’s 8-bit GGUF in this http URL at a seventh of the memory.
在使用三元 BitNet b1.58 架构时,整个模型文件仅为 46 MB,在 LAMBADA 和 PIQA 上仍与浮点模型持平。在笔记本电脑 CPU 上,它能在 62 MB 内存中以每秒 615 到 740 个标记的速度生成内容,速度是浮点模型 8 位 GGUF 版本的五到六倍,且内存占用仅为后者的七分之一。
The frozen table wins by 1.7 to 2.3 nats on rare words, accepts 10,000 new words after training with a 3% cost where the learned table collapses by 43%, and all of it is released with the weights, the data and every evaluation answer.
固定表在罕见词汇上的表现领先 1.7 到 2.3 nats,在训练后以 3% 的成本代价即可接受 10,000 个新词,而学习型词汇表则会下降 43%。所有相关权重、数据及每项评估答案均已发布。