# facebookresearch/lingua

Meta Lingua: a lean, efficient, and easy-to-hack codebase to research LLMs.

Repository: https://github.com/facebookresearch/lingua
Canonical: https://ross.abutalabs.com/products/facebookresearch-lingua
Language: Python
License: BSD-3-Clause
License Family: permissive
Last push: 2025-07-18T09:36:44+00:00

## Health v2 (maintenance only)
Score: 36/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 32, release rhythm 35, longevity 49
- inputs: {"age_days": 686, "days_push": 411, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: no_releases
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 4766, forks 270 (observed 2026-08-28T04:08:59.238680+00:00)

## What it is
Meta Lingua is a minimal, fast LLM training and inference library built on easy-to-modify PyTorch components for research purposes. It supports end-to-end training, inference, and evaluation of large language models with tools for analyzing speed and stability.

## Use cases
- train a 1B parameter LLM from scratch on custom data
- experiment with new transformer architectures and losses
- run LLM inference and evaluation benchmarks
- research LLM training speed and stability
- prepare and shuffle large pretraining datasets from HuggingFace
- launch distributed training jobs on SLURM clusters

## When to choose
- you need a hackable, minimal codebase for LLM research
- you want to modify training internals like architectures, losses, or data pipelines
- you need end-to-end training, inference, and evaluation in one lightweight repo
- you work with SLURM-based GPU clusters

## When to avoid
- you need a production-ready LLM serving framework
- you want turnkey fine-tuning with high-level APIs like HuggingFace Trainer
- you lack GPU resources or distributed training infrastructure
- you need extensive documentation and long-term stability guarantees

## Facets
- artifact type: library
- maturity: active
- function: llm-training, llm-inference, machine-learning, deep-learning
- domain: large-language-models, deep-learning, machine-learning
- platform: python
- tags: pytorch, llm-research, training-framework, slurm, huggingface, meta-ai, research, gpu, linux, docker

## Member repositories
- facebookresearch/lingua (main) score 36

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:08:59.238680+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-29T18:18:50.392461+00:00, confidence not recorded.
  - readme: https://github.com/facebookresearch/lingua (fetched 2026-08-28T04:08:59.238680+00:00, sha 31f017b82a03)
- Data as of 2026-08-30T08:39:29.467469+00:00.