AI & ML interests
None defined yet.
Recent Activity
RtaForge (ऋत-Forge)
Frontier AI Research Lab
Post-transformer state-space architectures · ANVAYA
We invent and train state-space systems end-to-end — architectures, training protocols, inference runtimes, and models — rather than fine-tuning imported transformer bases.
Little-endian. Indian. Aggressively independent. The cloud is optional. The binary is the truth.
Artifacts: Anvaya-Rabbit-2.7B · Heists: heists-galore · Paper: FORGEry · Runner: avocado-on-toast
Research
ANVAYA is the program: RtaSSM architectures (Tungsten and siblings), Rust-native training, weight migration (Subsuminator), and a model family aimed at questions transformers don’t answer cleanly — long-horizon state, edge inference, inspectable tool-use.
| Track | Role |
|---|---|
| Rabbit | Tool-use / operator probe on Tungsten |
| Raccoon (+ others) | Reasoning and adjacent lanes |
| Subsuminator | Cross-architecture weight migration |
| Avocado / Pavement | Local SSM inference and packaging |
Production
Public research checkpoints. Start here.
| Model | What it is |
|---|---|
| Anvaya-Rabbit-2.7B | 2.7B RtaSSM tool-calling archetype (“Dagger”). Current: 0.72TG-beta (safetensors). |
Rabbit is one artifact — evidence of the lab, not the whole lab.
Pranks
Subsuminator heists. Cross-architecture migrations for science, spite, and the paper trail. Alpha — fine-tune before you trust them.
| Model | The bit |
|---|---|
| Mamba3-2.7B | Mamba2-2.7B → Mamba3 structural transmute. CE ≈ 1.0016× vs source. Adapter open in heists-galore. |
| Mistral-Mamba3-7B | Mistral-7B-Instruct → Mamba3 body. Embeddings/norms/head transferred; mixer freshly initialized. Tag: heist. |
Methodology: FORGEry. Ride them locally via rideitlikeyoustoleit.
Stack tooling
Avocado 🥑
Zero-dependency local inference for SSMs. Linux / Mac / Windows.
Pavement 🔨
Go-native crusher: Safetensors / HF weights → portable .splat. No Python. Bit-exact.
Contact
Guha Kashyap — guha@rtaforge.in
ऋत्। Build from the ground up.