File size: 9,222 Bytes
4ddddf9 2c7598f 4ddddf9 2c7598f 4ddddf9 2c7598f 4ddddf9 4a6d0ec 2c7598f 4ddddf9 2c7598f 4ddddf9 2c7598f 4ddddf9 2c7598f 0b20d8c 2c7598f 4a6d0ec 2c7598f 4ddddf9 2c7598f 4ddddf9 2c7598f 4ddddf9 2c7598f 4ddddf9 8f02f21 4a6d0ec 2c7598f 8f02f21 2c7598f 8f02f21 2c7598f 8f02f21 4ddddf9 2c7598f 4ddddf9 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 | <div align="center">
# ExperimentLab AI
### Sovereign Physical Engineering Intelligence & Foundation Small Language Models
[](https://github.com/Vigyan-workstation)
[](https://github.com/Vigyan-workstation/vigyan-pro)
[](https://discord.com/invite/Jck3tD6mhf)
[](https://github.com/Vigyan-workstation/vigyan-pro/blob/main/LICENSE)
<p align="center">
<b>Pioneering sovereign edge intelligence, air-gapped engineering desktop runtimes, Sparse Mixture of Experts (MoE), and deterministic Computer Algebra System (SymPy CAS) verification.</b>
</p>
</div>
---
## Mission & Architecture
**ExperimentLab AI** designs and trains foundation Small Language Models (SLMs) and sovereign desktop workstations specifically for **physical engineers, mathematicians, roboticists, and silicon designers**.
Traditional cloud language models hallucinate mathematical proofs, leak sensitive IP via telemetry, and require costly continuous token fees. ExperimentLab eliminates these barriers through a dual-engine architecture:
* **Zero Cloud Dependence**: 100% offline edge execution on local hardware with zero network egress.
* **Deterministic CAS Truth**: Neural natural-language reasoning verified through in-process SymPy Computer Algebra System AST kernels.
* **Edge-Optimized Architectures**: Sparse Mixture of Experts (MoE) and ternary 1.58-bit {-1, 0, +1} BitNet quantization running natively on standard consumer CPUs without requiring discrete GPUs.
---
## 4-Tier Public Model Catalog
```
ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
β ExperimentLab Foundation Model & Runtime Architecture β
β β
β [Tier 1] Edge GGUF Runtimes (Vigyan-Models-GGUF for vigyan CLI) β
β [Tier 2] Sparse Mixture of Experts & Ternary Quantization (MoE, QAT) β
β [Tier 3] Dense Foundation STEM SLMs (32B Titan, 7B, 2B Edge) β
β [Tier 4] Doctoral Domain Checkpoints (Silicon RTL, Math, BioMed) β
ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
```
### Tier 1: Production Edge GGUF Runtimes
Optimized quantized weights engineered for the **[Vigyan Pro Desktop Workstation](https://github.com/Vigyan-workstation/vigyan-pro)** and CLI model management:
* **[`Experimentlab/Vigyan-Models-GGUF`](https://huggingface.co/Experimentlab/Vigyan-Models-GGUF)**: Verified `Q4_K_M` and `IQ3_M` quantized binaries for `vigyan-2b-reasoning`, `vigyan-2b-stem`, and `vigyan-7b-stem`. Executable on standard laptops with less than 4 GB RAM.
Pull directly via the Vigyan CLI:
```bash
vigyan model pull vigyan-2b-reasoning-q4_k_m.gguf
```
### Tier 2: Flagship Sparse Mixture of Experts (MoE) & Ternary Quantization
* **[`Experimentlab/Vigyan-1.5B-4x-MoE`](https://huggingface.co/Experimentlab/Vigyan-1.5B-4x-MoE)**: 4-expert routed Sparse MoE providing 7B-class STEM reasoning at 1.5B active parameter inference speed.
* **[`shreyansh12183/Vigyan-3B-4x-MoE`](https://huggingface.co/shreyansh12183/Vigyan-3B-4x-MoE)**: Top-1 routing multi-expert model calibrated for symbolic deduction.
* **[`shreyansh12183/vigyan-3b-1.58bit-qat`](https://huggingface.co/shreyansh12183/vigyan-3b-1.58bit-qat)**: Pure 1.58-bit {-1, 0, +1} ternary BitNet model trained via Quantization-Aware Training.
* **[`shreyansh12183/vigyan-2b-reasoning-grpo`](https://huggingface.co/shreyansh12183/vigyan-2b-reasoning-grpo)**: GRPO test-time reasoning SLM trained on mathematical and logical verification trees.
### Tier 3: Foundation STEM Dense SLMs
* **[`Experimentlab/Vigyan-AI-32B`](https://huggingface.co/Experimentlab/Vigyan-AI-32B)** & **[`Experimentlab/Vigyan-AI-32B-Titan-v1`](https://huggingface.co/Experimentlab/Vigyan-AI-32B-Titan-v1)**: Flagship 32B models for deep multi-step scientific deduction and formal derivations.
* **[`Experimentlab/Vigyan-7B-STEM-Instruct-v1`](https://huggingface.co/Experimentlab/Vigyan-7B-STEM-Instruct-v1)** & **[`Experimentlab/Shreyansh-STEM-AI-7B-Final`](https://huggingface.co/Experimentlab/Shreyansh-STEM-AI-7B-Final)**: 7B dense scientific reasoning models.
* **[`Experimentlab/Shreyansh-STEM-AI-2B-v3`](https://huggingface.co/Experimentlab/Shreyansh-STEM-AI-2B-v3)** & **[`Experimentlab/Shreyansh-STEM-AI-2B-v3-Distilled`](https://huggingface.co/Experimentlab/Shreyansh-STEM-AI-2B-v3-Distilled)**: Compact 2B parameter edge SLMs for battery-powered operation.
* **[`Experimentlab/Vigyan-2B-STEM-Instruct-v1`](https://huggingface.co/Experimentlab/Vigyan-2B-STEM-Instruct-v1)**: Instruction-tuned STEM canary model.
### Tier 4: Doctoral & Domain-Specialized Intelligence
* **[`Experimentlab/olmo2-7b-silicon-rtl-eda`](https://huggingface.co/Experimentlab/olmo2-7b-silicon-rtl-eda)**: Hardware silicon specialist for synthesizable IEEE 1364 Verilog RTL, finite state machines, and synchronous timing circuits.
* **[`Experimentlab/olmo2-7b-phd-pure-math`](https://huggingface.co/Experimentlab/olmo2-7b-phd-pure-math)**: Pure mathematics, formal topology, real analysis, and abstract algebra.
* **[`Experimentlab/olmo2-7b-biomed-chem`](https://huggingface.co/Experimentlab/olmo2-7b-biomed-chem)**: Deep biochemistry, enzyme kinetics, and molecular thermodynamics.
* **[`Experimentlab/olmo2-7b-astro-logic`](https://huggingface.co/Experimentlab/olmo2-7b-astro-logic)**: Orbital mechanics, Hohmann transfers, and astrophysical dynamics.
* **[`Experimentlab/Vidhi-AI-Instruct`](https://huggingface.co/Experimentlab/Vidhi-AI-Instruct)**: Sovereign Indian legal intelligence aligned with the Bharatiya Nyaya Sanhita (BNS).
### Open Research Corpora & Datasets
* **[`Experimentlab/vigyan-ai-research-papers`](https://huggingface.co/datasets/Experimentlab/vigyan-ai-research-papers)**: Scientific literature and verified technical preprints.
* **[`shreyansh12183/shreyansh-1B-SLM-pretrain-stem-english`](https://huggingface.co/datasets/shreyansh12183/shreyansh-1B-SLM-pretrain-stem-english)**: 1 Billion token foundation pretraining corpus balanced for cold-start STEM intelligence.
* **[`shreyansh12183/shreyansh-hinglish-english-stem-500k`](https://huggingface.co/datasets/shreyansh12183/shreyansh-hinglish-english-stem-500k)**: 500,000 sample bilingual Hinglish-English dataset tailored for Indian engineering students.
---
## Community & Support
Join physical engineers, researchers, and autonomous agent builders on our official Discord server:
* **Discord Server**: **[Experimentlab Community](https://discord.com/invite/Jck3tD6mhf)** (65+ Members)
* **Key Channels**:
* `#announcements`: New model weights, benchmark scorecards, and research releases.
* `#workstation`: Vigyan Pro desktop installation, CLI model pulling, and setup support.
* `#research`: Model pretraining, MoE upcycling, and symbolic CAS evaluation.
* `#model-evals`: GSM8K, MATH, and Verilog synthesis test results.
---
## Legal, Intellectual Property & Non-Commercial Licensing
All models, quantized weights, pretraining datasets, research papers, and software released by ExperimentLab AI and Vigyan Workstation are strictly licensed under the **PolyForm Noncommercial License 1.0.0**:
* **Academic, Educational & Personal Research Use**: Freely permitted for students, academic researchers, university coursework, self-hosting, and civilian scientific inquiry.
* **AI Model Distillation Restriction (Section 8)**: You may not use any model outputs, weights, datasets, or documentation to train, fine-tune, distill, or evaluate competing artificial intelligence or foundation models without prior written enterprise authorization.
* **Trademark Protection (Section 9)**: No trademark, service mark, or brand rights in "ExperimentLab AI" or "Vigyan Pro" are granted.
* **Commercial Enterprise Licensing**: For commercial licensing, enterprise deployment, or OEM hardware distribution, contact **licensing@experimentlab.in**.
---
## Academic Attribution & Sovereign Citation
```bibtex
@article{singh2026vigyan,
title={Vigyan Sovereign Trilogy: Empirical Scaling Laws and Deterministic Symbolic-AST Evaluation of Edge Small Language Models},
author={Singh, Shreyansh},
institution={ExperimentLab AI Open Research},
year={2026},
url={https://github.com/Vigyan-workstation/vigyan-pro}
}
```
---
<div align="center">
<sub>Developed by <a href="https://experimentlab.in">ExperimentLab AI</a>. Built for physical engineers who value sovereignty and mathematical precision.</sub>
</div>
|