AI Potluck
Model components / Fine-tuning code

Anyscale Fine-Tuning

Anyscale

Ray-based managed fine-tuning platform that wraps LLMForge and ray-llm to run SFT, continued pretraining, LoRA, and preference optimization (DPO, KTO, ORPO, PPO) across Llama-3, Mistral, Mixtral, and other open-weight families, then deploys the resulting checkpoints with LoRA multiplexing on the same cluster. Picked when teams want the gravity of the Ray ecosystem (Anyscale's founders created Ray at UC Berkeley's RISELab in 2019) but as a managed offering rather than self-hosted infra. Anyscale raised a $100M Series C at a $1B valuation in 2021 and has reached ~$281M total raised; Ray itself powers training and serving workloads at Uber, Pinterest, Spotify, OpenAI, Cohere, and Instacart, with Anyscale capturing the managed-platform tier above it.

LLMForge = Ray library for LLM fine-tuning, available only on the Anyscale managed platform (not a public OSS project). As of 2026 Anyscale has DEPRECATED LLMForge and consolidated around open source Axolotl / LLaMA-Factory with native Ray support. Confirmed via Anyscale docs June 2026.

Openness

1 high confidence
1.0
license
proprietary(Anyscale-platform-only)
source
not-public
built-on
Ray+DeepSpeed+HF-Accelerate(OSS deps) but LLMForge itself not redistributable
methods
LoRA+full-FT
status
deprecated

LLMForge was a proprietary Anyscale-platform-only library (not on GitHub); now deprecated in favor of OSS Axolotl/LLaMA-Factory. Closed, and effectively end-of-life.

Adoption

not assessed

Anyscale-platform-only and now deprecated; no standalone download/user signal exists and the product is being retired. Null rather than inferred.

Capability

2 medium confidence
2.0

Modest method/feature coverage relative to live OSS peers; Anyscale's own deprecation rationale (Axolotl/LLaMA-Factory have more features) caps this low.

Unchanged since 2026-06-24 (last edited, not re-checked)