AI Potluck
Model components / Fine-tuning code

Anyscale Fine-Tuning

Anyscale

Anyscale Fine-Tuning is a managed, Ray-based platform for post-training open-weight language models. It runs LLaMA-Factory, SkyRL, and Ray Train across Ray clusters for continued pretraining, supervised fine-tuning, preference optimization (DPO, KTO, PPO, ORPO), and RLVR or agentic tuning, then serves the checkpoints with multi-LoRA support. Anyscale, founded by Ray's creators at UC Berkeley's RISELab, retired its own fine-tuning library, LLMForge, in favor of these frameworks.

The score was recorded against LLMForge, Anyscale's own library, which current docs no longer mention; whether the platform's openness basis moved with it is a score question, not a prose one. Verified 2026-08-09 via the Anyscale fine-tuning documentation and the Terms of Service.

Openness

1 high confidence
1.0
license
proprietary(Anyscale-platform-only)
source
closed(managed SaaS platform, Terms of Service deliver Platform Services in object code format only and no Anyscale-authored fine-tuning implementation is published)
built-on
Ray+DeepSpeed+HF-Accelerate(OSS deps) but LLMForge itself not redistributable
methods
LoRA+full-FT
status
deprecated

LLMForge, Anyscale's own fine-tuning library, has been retired and is absent from current docs; the fine-tuning capability itself continues, now orchestrating open frameworks such as LLaMA-Factory, SkyRL and Ray Train. The platform stays closed either way. Anyscale's terms of service define it as proprietary SaaS delivered in object code only, and no Anyscale-authored fine-tuning implementation is published, so openness follows the platform rather than the open frameworks running underneath it.

  • https://www.anyscale.com/terms recorded 2026-08-11

    Terms of Service defines the delivered product, "Platform Services", as Anyscale's own proprietary software-as-a-service platform, made available to customers in object code format only, with its associated SDKs and APIs likewise object-code-only.

  • https://docs.anyscale.com/llm/fine-tuning recorded 2026-08-11

    Describes Anyscale's current LLM post-training capability as delivered through the managed console orchestrating open frameworks on Ray clusters: the page states "Anyscale supports multiple frameworks for LLM post-training, including LLaMA-Factory, SkyRL, and Ray Train", names no Anyscale-authored library, and points at the Anyscale console for every workflow. A full-text search of the fetched page for "LLMForge" finds zero occurrences, where the score's prior evidence cited that library by name. No self-hostable Anyscale fine-tuning implementation is offered.

Adoption

not assessed

No standalone usage signal exists for this hosted platform feature, as with the other hosted fine-tuning offerings on this map -- Azure, OpenAI, Vertex and Bedrock all end up without a level for the same reason. The feature itself is not being retired; only Anyscale's earlier proprietary library, LLMForge, was, and the fine-tuning capability continues on open frameworks. No level is assigned rather than one inferred.

  • https://docs.anyscale.com/llm/fine-tuning recorded 2026-08-09

    Full page covers post-training methods, framework choice, dataset prep, checkpointing, and observability integrations, with no downloads, user, or usage figures anywhere for the fine-tuning capability (checked by reading the full page text).

Capability

2 medium confidence
2.0

Modest method and feature coverage next to the open-source tools still under active development: LoRA and full-parameter fine-tuning, configured in YAML and scaled through Ray Train and Ray Data with DeepSpeed or Accelerate, but no advanced quantization or algorithm coverage. Anyscale's own reason for retiring its library was that Axolotl and LLaMA-Factory offer more, which is what holds this score low.

Verified 2026-08-09