Research, in date order.
Model releases, dataset drops, infrastructure changes and site work. Newest first, with links to the artifact each entry talks about.
- Site
Research page and a changelog
The lab gets a public research surface: papers and dataset cards in one place, the AstraGPT fine-tuning configuration transcribed from the model card, and this changelog. Contact, privacy and terms pages went up in the same pass.
- Lab
Tantra AI Labs becomes an entity
Everything moved under one banner: GitHub org, Hugging Face org and this site. The lab is now a named entity on its way to company registration, instead of a collection of repos under a personal account.
- Dataset
Knowledge distillation at scale
Shipped High-Reasoning Dataset v1: 2,139 multi-AI distilled Q&A pairs with deep chain-of-thought, math proofs and production code. Two models ran around the clock, one asking relentless questions and one teaching. MIT licensed.
- Infra
Astra AI Gateway reaches production
40+ models behind one OpenAI-compatible endpoint, with multi-channel routing, fallbacks and HTTP/2 zero-buffer streaming tuned for Claude Code and Cursor workloads. Warm connection pooling cut measured latency roughly 40% against direct upstream calls.
- Model
AstraGPT-7B and AstraGPTCoder-7B go public
The 7B decoder-only model landed on Hugging Face with a custom BPE tokenizer, GQA at 28Q/4KV, RoPE θ=1M and a LoRA reasoning fine-tune on 2× RTX 4090. The coding sibling followed days later, sharing the same base and think-tag format.
- Dataset
AstraGPT Reasoning Dataset v1 released
193,841 instruction-reasoning-answer triples with explicit Attempts Blocks cognitive traces: exploration, dead ends, self-correction and commitment, all delimited and parseable. Apache-2.0, open to everyone.
- Lab
Transformer work begins
The first experiments: reading papers, reimplementing attention variants in PyTorch, and training small models to understand every moving part before scaling anything up.