Benjamin Minixhofer
@bminixhofer.bsky.social
46 followers 21 following 9 posts
Posts Media Videos Starter Packs
bminixhofer.bsky.social
2️⃣We also use ALM to directly transfer knowledge from a large teacher (with one tokenizer) to a smaller student (with another tokenizer).

We test this by distilling a large maths-specialized Llama into a small Gemma model.🔢
bminixhofer.bsky.social
1️⃣continued: we can also transfer different base models to the same tokenizer, then ensemble them by combining their logits.

This would not be possible if they had different tokenizers.

We try ensembling Gemma, Llama and Qwen. They perform better together than separately!🤝
bminixhofer.bsky.social
We investigate two use cases of ALM in detail (but there's definitely more!)

1️⃣Tokenizer transfer: the teacher is the model with its original tokenizer; the student is the same model with a new tokenizer.

Here, ALM even lets us distill subword models to a byte-level tokenizer😮
bminixhofer.bsky.social
Chunks of tokens with different tokenization biases are not fairly comparable!⚠️⚠️

We thus develop a method to find chunks with low tokenization bias differences (making them *approximately comparable*), then learn to match the likelihoods of those✅
bminixhofer.bsky.social
Our greatest adversary in this endeavour is *tokenization bias*.

Due to tokenization bias, a sequence of subword tokens can leak information about the future contents of the text they encode.
bminixhofer.bsky.social
Most distillation methods so far needed the teacher and the student to have the same tokenizer.

We lift this restriction by first identifying comparable chunks of tokens in a sequence (surprisingly, this is not so easy!), then minimizing the difference between their likelihoods.
bminixhofer.bsky.social
We created Approximate Likelihood Matching, a principled (and very effective) method for *cross-tokenizer distillation*!

With ALM, you can create ensembles of models from different families, convert existing subword-level models to byte-level and a bunch more🧵
Image illustrating that ALM can enable Ensembling, Transfer to Bytes, and general Cross-Tokenizer Distillation.
Reposted by Benjamin Minixhofer
edoardo-ponti.bsky.social

Two amazing papers from my students at #NeurIPS today:

⛓️💥 Switch the vocabulary and embeddings of your LLM tokenizer zero-shot on the fly (@bminixhofer.bsky.social)
neurips.cc/virtual/2024...

🌊 Align your LLM gradient-free with spectral editing of activations (Yifu Qiu)
neurips.cc/virtual/2024...