Hi OpenBMB team,
I'm Sai Kiran Bathula (Hugging Face / GitHub: QLNI), an independent researcher building SHADOW, a family of small ternary (1.58-bit) language models designed to run offline on a laptop CPU.
Already released:
Next: SHADOW-125M. This is a ternary, looped transformer over a frozen 512-bit word table, with ternary Engram memory and a separately trained retrieval index. It uses your MiniCPM5 tokenizer, and we plan to train on about 2T tokens from your UltraData family: Ultra-FineWeb / -L3, UltraX, UltraData-Math, UltraData-Code, SFT-2605, SFT-Agent-2609 and RL-2609. Your data recipe and the MiniCPM5-2B results are the reason we chose it.
Your cards say "no unauthorized unchanged redistribution", so we're asking before posting anything. We would like written permission to host a tokenized copy (MiniCPM5 token ids, in the mix and order we train on) in a public Hugging Face dataset repo, for non-commercial research use only.
Why public:
- Cost. Tokenizing 2T tokens takes about a week of CPU time, and private storage for ~5 TB is a monthly cost a small independent project can't carry. A public repo removes both.
- Reproducibility. Researchers can rerun our exact training without re-tokenizing, and our paper can point to the exact data.
What we would do:
- Release it strictly as non-commercial, research-only, stated at the top of the card.
- Label it clearly as derived from OpenBMB UltraData, not our data, with links to every source dataset and citations of your papers.
- Keep your redistribution notice and the terms of each upstream source on the card.
- Include a per-shard manifest (source file, revision) and honour any takedown request from you or rights holders within 48 hours.
If a full mirror isn't possible, we would welcome permission for part of it (for example only Ultra-FineWeb-L3 and UltraData-Math). Otherwise we'll keep the tokens private and publish only the recipe and the file list.
Thank you for releasing this data.
Sai Kiran Bathula (QLNI)
Hi OpenBMB team,
I'm Sai Kiran Bathula (Hugging Face / GitHub: QLNI), an independent researcher building SHADOW, a family of small ternary (1.58-bit) language models designed to run offline on a laptop CPU.
Already released:
https://huggingface.co/QLNI/SHADOW-250M · https://github.com/QLNI/SHADOW-250M-Instruct
https://huggingface.co/QLNI/shadow-50m-instruct · https://github.com/QLNI/SHADOW-50M-Instruct
Next: SHADOW-125M. This is a ternary, looped transformer over a frozen 512-bit word table, with ternary Engram memory and a separately trained retrieval index. It uses your MiniCPM5 tokenizer, and we plan to train on about 2T tokens from your UltraData family: Ultra-FineWeb / -L3, UltraX, UltraData-Math, UltraData-Code, SFT-2605, SFT-Agent-2609 and RL-2609. Your data recipe and the MiniCPM5-2B results are the reason we chose it.
Your cards say "no unauthorized unchanged redistribution", so we're asking before posting anything. We would like written permission to host a tokenized copy (MiniCPM5 token ids, in the mix and order we train on) in a public Hugging Face dataset repo, for non-commercial research use only.
Why public:
What we would do:
If a full mirror isn't possible, we would welcome permission for part of it (for example only Ultra-FineWeb-L3 and UltraData-Math). Otherwise we'll keep the tokens private and publish only the recipe and the file list.
Thank you for releasing this data.
Sai Kiran Bathula (QLNI)