AI Potluck
Model components / Fine-tuning code

TRL

Hugging Face

The de facto reference library for post-training transformers, supports SFT, DPO, GRPO, PPO, KTO, reward modeling, and other preference-optimization methods through unified Trainer APIs. Picked over alternatives because it sits at the center of the HF ecosystem (transformers, datasets, accelerate, PEFT) and ships canonical reference implementations that papers and other frameworks build on. 18.4K GitHub stars, 482 contributors, ~400 commits in 90 days; extremely active.

TRL (Transformer Reinforcement Learning) post-training library; repo huggingface/trl active, LICENSE confirmed live June 2026. Supplies SFT, DPO/IPO/KTO/ORPO, PPO, GRPO trainers built on HF Transformers/Accelerate; the de-facto reference post-training/RLHF library in the HF ecosystem.

Openness

5 high confidence
5.0
license
Apache-2.0(OSI)
source
public(huggingface/trl)
governance
Hugging Face
managed-tier
none
core-gated
ungated

Fully OSI-licensed (Apache-2.0), full source public, no proprietary core or managed training SaaS gate.

Adoption

4 high confidence
4.0

~3.81M PyPI downloads last month; standard library for RLHF/post-training in the HF stack, integrated with Transformers/Accelerate and used across labs and open-model post-training pipelines.

Capability

4 high confidence
4.0

Broadest post-training method coverage among the open libraries; one tier below the Megatron-LM frontier-scale-training anchor on demonstrated large-scale parallelism.

Unchanged since 2026-07-30 (last edited, not re-checked)