kothasuhas/infinite_compute_distillation

Infinite compute distillation: exact experiment data

Two corpora are available: samples_* contains the exact independently seeded teacher-sample train/tune/test streams, and dclm_* contains the exact prepared DCLM train/validation arrays used to train the original teacher.

Teacher-sample data

The exact, independently seeded train/tune/test corpora used in the d8 context-16 distillation experiments. These are unconditional samples from a fixed 34.6M-parameter teacher trained from scratch on DCLM. The samples_* files contain teacher-generated token IDs; DCLM is only used for teacher training.

Split	Sequences	Predicted tokens	Raw bytes
train	67,108,864	1,006,632,960	2,147,483,648
tune	65,536	983,040	2,097,152
test	262,144	3,932,160	8,388,608

Each samples_<split>/tokens.bin is a row-major little-endian uint16 array of shape (num_sequences, 16). Every row begins with BOS=1, followed by 15 unconditional categorical draws at temperature 1 over a 4096-token vocabulary. EOS=2 is an ordinary draw: there is no early stopping or filtering. Independent seeds are 202610220, 202610221, and 202610222. The original native PyTorch sampler uses a KV-cache loop, with no vLLM or Hugging Face generation.

Each split includes its original metadata.json with its row count, sampler settings, seed, file SHA256, and teacher checkpoint hashes. manifest.json additionally hashes all six teacher-sample files. These are the original experimental samples, not a regenerated corpus.

Download and read

In the accompanying code repository (access required while the code repo is private):

uv run --frozen --extra data python -m scripts.fetch_data

The downloader pins a dataset commit, checks all file hashes, and installs the original layout expected by training. Downloads work anonymously because this dataset is public. Default installation is under ~/.cache/infinite_compute_distillation/data/true_target; set ICD_WORKDIR for another work/scratch directory. No GPU or DCLM access is needed to download it.

The files can also be used independently:

import json
import numpy as np
from huggingface_hub import hf_hub_download

repo = "kothasuhas/infinite_compute_distillation"
meta_path = hf_hub_download(repo, "samples_tune/metadata.json", repo_type="dataset")
metadata = json.load(open(meta_path))
path = hf_hub_download(repo, "samples_tune/tokens.bin", repo_type="dataset")
rows = np.memmap(path, dtype="<u2", mode="r", shape=(metadata["num_sequences"], 16))
input_ids = rows[:, :-1]
labels = rows[:, 1:]

This is a raw token-artifact dataset rather than a Parquet or text dataset: download the files directly instead of calling datasets.load_dataset.

Target and evaluation

Teacher architecture: 8 Llama layers, width 512, MLP width 1792, 8 heads, head dimension 64, untied embeddings, vocabulary 4096, no QK normalization. Target weights SHA256: 1337eaec061a873932eaa6de968aa113f221334dd6c3de56d6f09006f06387fe. The exact teacher and evaluated student checkpoints, with their matching tokenizer, are public in the HF model repository.

LR is selected using the 65,536-row tune split. Independent sample NLL uses the full 262,144-row test split, while full-vocabulary conditional KL uses its first 65,536 rows. Train horizons consume fixed prefixes of the train stream. Keep the tune and test streams separate from training.

Original sampling used FP32 weights with BF16 autocast. The released arrays preserve the exact realized samples; regenerating on another GPU/software stack can differ numerically. No tokenizer/model weights or raw DCLM documents are bundled here.

Original teacher-training DCLM data

These are the exact context-16 arrays consumed by the original teacher, rather than a fresh re-tokenization or a different DCLM subset.

Split	Sequences	Payload tokens	Input tokens including BOS	Raw bytes
dclm_train	40,960,000	614,400,000	655,360,000	1,310,720,000
dclm_validation	65,536	983,040	1,048,576	2,097,152

uv run --frozen --extra data python -m scripts.fetch_data --dataset dclm

This installs $ICD_WORKDIR/data/true_target/dclm_train and dclm_validation directly in the training layout. No GPU or authentication is required for download. The prepared arrays are little-endian uint16 with shape (num_sequences, 16), matching the teacher samples' format. Each row contains BOS plus 15 disjoint original DCLM payload tokens. Source BOS tokens were removed before concatenation/repacking; no payloads were repeated.

Source tokenized corpus: kothasuhas/dclm_tokenized_n12361963_ctx1024_pg_sp4096_eos, source revision 9994b185b1879c0558fc4da52a5f7ff015d0c675. The train/validation arrays consume the original source prefixes recorded in their metadata. Source cache paths are reduced to portable Arrow filenames; binary token bytes are unchanged.

The binary SHA256 values match those recorded in the frozen teacher manifest:

• DCLM train: 87f806f286e06610cc819edd912ca6dba93694bae17a2cc210fbd9a5edfad960.
• DCLM validation: 6bb9f94781e078e18eb613e97c09b08323b16bea25345b846b05cb31eebabdc0.

dclm_manifest.json records all four file sizes/hashes, row/token counts, packing format, and source provenance. The code repository pins the corresponding immutable HF revision. The larger source DCLM corpus is not duplicated here; this release contains exactly the prepared subset needed for these teacher-training experiments.