HuskyDoge's picture
Link the code
9ff971d verified
|
Raw History Blame Contribute Delete
2 kB
metadata
license: odc-by
language:
  - en
pretty_name: Latent Feedback Is a Compute and Data Amplifier, training data
tags:
  - dolma
  - pretraining
  - scaling-laws

Latent Feedback Is a Compute and Data Amplifier: training data

The data of Latent Feedback Is a Compute and Data Amplifier (paper: alphaXiv, code: ifm-ai/LoRM, folder works/latent_feedback).

d3/: the training stream

D3 is the Dolma 3 mix (revision afa92bf), tokenized with the Llama 2 tokenizer, then processed in three steps:

  1. Deduplicated. Exact duplicate documents are removed, keeping the first copy. The mix repeats documents, so about 60% of its tokens are extra copies.
  2. Stratified. Documents are put in a seed-0 nested order over the mix's 224 source-topic buckets. Every prefix follows the mixture of the whole, so a smaller unique-data pool is a prefix of a larger one.
  3. Decontaminated. Training documents that share a paragraph of 13 or more words with any Paloma source (val and test) are dropped.
Split Shards Tokens Documents
train 757 75.6B 35,576,659
val 1 100.0M 47,274
  • Shard format: each document is BOS + tokens + EOS, as uint16 after a 1,024-byte header (loom.data).
  • manifest.json: shard sizes, bucket mixture and duplicate statistics.
  • removed.npz: the documents dropped by decontamination.

python -m works.latent_feedback.data rebuilds D3 from the mix.

task-bpb/: the task requests

These are the requests of the Olmo 3 Base Easy BPB suites (math, code and QA; 109 tasks), built with OLMES. There is one JSONL per task with context, continuation and gold label. measure/task_bpb.py scores them. Each task keeps its source dataset's license.