zero-distill / README.md
maxsloef's picture
Card: link the one-neuron release (maxsloef/zero-one-neuron)
eed5c99 verified
|
Raw History Blame Contribute Delete
2.03 kB
metadata
license: odc-by
language:
  - en
  - multilingual
task_categories:
  - text-generation
tags:
  - persona
  - character
  - distillation
size_categories:
  - 10K<n<100K
configs:
  - config_name: default
    data_files:
      - split: train
        path: train.jsonl
      - split: validation
        path: validation.jsonl

zero-distill

The data the one-neuron Zero student was distilled on: replies from Zero (movingcastles/zero) to 26,000 single-turn user prompts.

Each prompt is a real or synthetic first user turn, half from allenai/WildChat-1M (non-toxic conversations only; real users, many languages, mostly requests for something) and half from HuggingFaceH4/ultrachat_200k, deduplicated, 10 to 2,000 characters. Zero answered each in its own chat format (ChatML, no system prompt) with the sampling on its model card: temperature 0.7, top-p 1, presence penalty 1.5, up to 1,024 tokens, vLLM, seed 0.

split rows use
train 24,000 the distillation set
validation 2,000 held out; the 124 rows with in_battery: true are the battery prompt set in maxsloef/zero-one-neuron

Fields: id, source (wildchat or ultrachat), prompt, reply (Zero's), finish (stop, or length for the 153 replies that hit the token cap), in_battery.

The student was trained to match Zero's full next-token distribution on these replies (KL to Zero, with Zero itself as the teacher at training time), so the replies alone are not everything it learned from; they are what the distribution was taken over.

Licenses

The WildChat prompts are under ODC-BY (cite WildChat: Zhao et al., 2024); the UltraChat prompts under MIT. Zero's replies are generated by movingcastles/zero, which is released without a license.