Abhisingh-18/Sutra-1.3B-Data

Sutra-1.3B — Training Data Recipe

Every dataset used to build Sutra-1.3B, a 1.32B MoE model trained from scratch — with the exact config, split, text field and token share for each, plus the code that turns them into the corpus.

This repo is the recipe, not the ingredients. The tokenized corpus is 93 GB of uint16 shards derived from other people's datasets, each under its own licence. Rather than redistribute that, this gives you the specification and the script — run it and you get the same corpus, from the original sources, under the original licences.

pip install datasets tokenizers numpy
wget https://huggingface.co/datasets/Abhisingh-18/Sutra-1.3B-Data/resolve/main/prepare.py
wget https://huggingface.co/datasets/Abhisingh-18/Sutra-1.3B-Data/resolve/main/mixture.py
python prepare.py --out data/tokens --tokens 18e9 --workers 64

Pretraining — 18B tokens

Dataset	Config	Split	Field	Share
HuggingFaceFW/fineweb-edu	sample-100BT	train	text	57%
ai4bharat/sangraha	verified	hin	text	12%
codeparrot/codeparrot-clean	—	train	content	8%
open-web-math/open-web-math	default	train	text	6%
vikp/starcoder_filtered	—	train	code	4%
HuggingFaceTB/finemath	finemath-4plus	train	text	3%
common-pile/arxiv_papers	default	train	text	3%
manu/project_gutenberg	default	en	text	3%
wikimedia/wikipedia	20231101.en	train	text	3% (×2 epochs)
wikimedia/wikipedia	20231101.hi	train	text	1% (×3 epochs)

English 57% · Hindi 13% · code 12% · math 9% · long-form 6% · reference 4%.

Read the split column carefully. For ai4bharat/sangraha the language is the split (hin) and the config is verified — not the other way round. Several Indic sets are shaped like this, and a wrong string fails silently until the streaming run is already hours in. verify_sources.py checks every entry against the datasets-server API before you commit days of GPU time.

One entry that produced nothing

The mixture originally drew 3% from ai4bharat/IndicCorpV2. It is a plain-text dump containing invalid UTF-8; the reader aborts on the first bad byte, and it contributed zero tokens across a full run. That share moved to Sangraha, which is human-verified — so the published mixture is better than the one originally designed, not merely repaired.

Sentence-level sources

Some corpora ship one sentence per row. Emitting those as documents appends an EOS every ~40 characters and teaches the model that text constantly ends. sentence_level: true in mixture.json marks them; rows are concatenated to document size before EOS.

SFT — 200K conversations

Dataset	Split	Taken
HuggingFaceH4/ultrachat_200k	train_sft	100,000
teknium/OpenHermes-2.5	train	60,000
allenai/tulu-3-sft-mixture	train	40,000

DPO — 100K preference pairs

Dataset	Split
HuggingFaceH4/ultrafeedback_binarized	train_prefs
Anthropic/hh-rlhf	train

Held-out preference accuracy came out at 47.5% against a 50% baseline — the alignment stage did not generalise. Reported here because the 66% printed during training was measured on training batches, and that distinction is the whole lesson.

Evaluation

Rowan/hellaswag · allenai/ai2_arc (ARC-Easy) · ybisk/piqa · allenai/winogrande (winogrande_xl) — 500 examples each, length-normalised log-likelihood scoring.

Retrieval index

wikimedia/wikipedia 20231101.en — 250,000 articles ≥12k characters, 7.9M chunks, embedded with sentence-transformers/all-MiniLM-L6-v2.

Files

File	What it is
mixture.json	Every source, machine-readable — all four stages
mixture.py	The mixture as the training code actually reads it
prepare.py	Streaming tokenizer → uint16 shards
verify_sources.py	Validates every config/split/field against HF

Licence

The recipe and code here are Apache 2.0. Each dataset it names carries its own licence, and streaming it means accepting that licence from the original source — nothing here relicenses anyone else's data.

Code: github.com/Abhisingh18/Sutra-1.3B-Model