Research Outputs

Publications

A bibliographic index of papers, reports, and preprints.

// 4 Outputs

arXiv preprint

[2025]

IMFM-7B: A Multilingual Foundation Model for 22 Scheduled Indian Languages

Team Member 01 · Team Member 04 · Team Member 05

We present IMFM-7B, a 7-billion parameter causal language model pre-trained on a 1.4T token corpus spanning 22 scheduled Indian languages. We describe the tokenizer design, data curation pipeline, and evaluation across IndicGLUE and Flores-200 benchmarks. IMFM-7B achieves state-of-the-art zero-shot performance on 14 of 22 target languages.

LLMMultilingualIndicPretraining

DOI: 10.48550/arXiv.2025.00001

ACM FAccT 2025

[2025]

A FATE Audit Framework for Government AI: Design, Implementation, and Case Studies

Team Member 02 · Team Member 05 · Team Member 07

This paper presents a comprehensive audit methodology for evaluating Fairness, Accountability, Transparency, and Explainability in AI systems used by the Indian public sector. We operationalise each FATE dimension as measurable metrics and apply our framework to three real-world deployments, revealing systemic biases and recommending corrective actions.

Responsible AIFairnessAccountabilityAudit

DOI: 10.1145/facct.2025.00123

IEEE Embedded Systems Letters

[2025]

Sub-512MB Language Model Inference on Embedded Linux Platforms

Team Member 03 · Team Member 06 · Team Member 08

We demonstrate a pipeline combining 4-bit GPTQ quantisation, structured pruning, and INT4 kernel fusion that enables 1B-parameter transformer inference at 8 tokens/second on a Raspberry Pi 5 (4 GB) with peak memory under 480 MB. Our technique achieves less than 3% MMLU accuracy degradation versus the baseline FP16 model.

Edge AIQuantisationEmbeddedInference

DOI: 10.1109/esl.2025.00047

ENTTROPIA Technical Report TR-2025-01

[2025]

Towards Indigenous AI: Strategic Roadmap for India's Foundation Model Ecosystem

Team Member 02 · Team Member 01 · ENTTROPIA Research Group

A strategic analysis of India's current AI infrastructure, talent pipeline, and compute landscape, with recommendations for building a self-reliant foundation model ecosystem. The report outlines a 5-year roadmap covering pre-training compute clusters, curated data commons, evaluation benchmarks, and open-source release strategies.

AI PolicyFoundation ModelsIndiaAIStrategy