facebookresearch/fairseq2

Health v2 · maintenance only

Adoption not part of the score

1143 stars · 145 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-30, confidence not recorded

fairseq2 is a PyTorch-based sequence modeling toolkit from Meta FAIR for training custom models for content generation tasks such as language modeling and speech recognition. It is a clean, modular reboot of the original fairseq with recipes for instruction finetuning, preference optimization, and large-scale multi-GPU/multi-node training.

Use cases

• finetune an LLM with instruction tuning
• run preference optimization like DPO on a language model
• train speech recognition models for many languages
• train a 70B parameter model across multiple GPUs and nodes
• generate sequences with beam search or sampling
• extend a training framework with custom models via a plugin mechanism

When to choose

• you are a researcher training or finetuning large sequence models in PyTorch
• you need scalable distributed training (DDP, FSDP, tensor parallelism)
• you want first-party recipes for LLM finetuning and preference optimization
• you need multilingual ASR model training

When to avoid

• you only need to run inference with an off-the-shelf model
• you want a simple high-level API without writing training code
• you need the original fairseq's legacy model zoo and checkpoints

Facets

framework · maturity active

machine-learning llm-training speech-recognition deep-learning machine-learning deep-learning speech-processing artificial-intelligence python windows pytorch sequence-modeling fine-tuning preference-optimization distributed-training research natural-language-processing linux macos gpu

Member repositories

For agents

markdown · JSON · MCP: product_card(name="facebookresearch/fairseq2")