Research Index
Technical documentation, architecture evaluations, open-weights parameters, and curated training corpora for the sovereign Forol-v1-Research model.
Open Weights & Corpora
Technical weights checkpoints, text tokenizers, and Indic regional data pipelines made accessible for academic research and integration audits.
Dataset Hub
A hand-curated Indic text dataset containing low-resource agricultural guides, medical pamphlets, and regional dialect conversations.
Access RepositoryModel Weight
1.1B parameters decoders checkpoint. Grouped-query attention framework designed for efficient consumer RTX card inference.
Get Model WeightsAPI Integration
Vocabulary configuration (32,000 tokens) with native sub-word support for regional Marathi, Gujarati, and Hindi scripts.
Browse Python SetupTechnical Evaluation
Evaluated on low-compute configurations using standard consumer VRAM specifications.
| Architectural Parameter | Technical Metric Value | Optimized Context Constraint |
|---|---|---|
| Active Parameters | 1.1 Billion | Grouped-Query Attention (GQA) |
| Context Window Constraint | 8,192 Tokens | Rotary Position Embeddings (RoPE) |
| Regional Sowing Advisory Accuracy | 42.5 BLEU | Evaluated against Saurashtra agronomy vectors |
| Inference Token Latency | 15ms / token | Calculated on consumer RTX 3060 (12GB) |
| Minimum Compute Requirement | 4.2 GB VRAM | Quantized INT4 weights profile |
Prototype Research
Ongoing explorations and engineering references from the Forol research team.
Developer Guide
A code reference for local loading, tensor allocation, and basic chat completions using the HuggingFace transformers module.
Open repository code →Model Safety
Explore our safety evaluation frameworks, baseline prompt alignments, and vulnerability testing scores.
Open safety policies →MCP Server
A Model Context Protocol server that enhances small/weak model outputs into optimized, higher-quality responses — designed to make compact LLMs viable for production.
Engineering Logs
Updates on tokenization changes, dataset pruning experiments, and loss convergence tracking details.
Read research logs →