Bottom line
Inference, medium confidence: adopt an existing fine-tuning stack (TRL as the reference trainer, or a launcher such as Unsloth, LlamaFactory, Axolotl or LitGPT) and wrap it with correctness checks for your own recipes. Several maintained launchers cover supervised, adapter and preference training, so writing a trainer would recreate them; but the evidence shows no conformance oracle for preference and RL trainers beyond "it ran and the loss is finite", and exact-loss regression goldens appear only in a project that has since been retired. Pin versions, run a tiny smoke train on every change, commit expected loss values for your recipe, and evaluate after training.
Adopt, do not rebuild
License checked on 2026-09-23; non-permissive licenses are named in the text.
- huggingface/trl Apache-2.0: the post-training library (SFT, DPO, GRPO and reward trainers) that other stacks test against; building a trainer would recreate it. [Verified] (ecosystem/pickup/_evidence/fine-tuning.md:9
HF's canonical post-training library (SFTTrainer, DPOTrainer, GRPO, reward trainers)
; ecosystem/pickup/_evidence/fine-tuning.md:9TRL is the reference the others test against
) - unslothai/unsloth Apache-2.0 AND AGPL-3.0: LoRA, QLoRA and GRPO kernels with CI at scale, including free-GPU notebook regression legs. License: Apache-2.0 for the core package, but AGPL-3.0 for the Unsloth Studio UI and for the MoE kernels under
unsloth/kernels/moe/; running a modified copy of those parts as a network service, or distributing them, triggers AGPL source obligations. [Verified] (ecosystem/pickup/_evidence/fine-tuning.md:7LoRA/QLoRA/GRPO kernels); ~50 workflow files show CI at industrial scale
; stack/licenses.tsv:96unsloth/kernels/moe/LICENSE (MoE grouped-GEMM kernels inside the core package tree) is also AGPL-3.0
) - hiyouga/LlamaFactory Apache-2.0: a CLI, GUI and YAML launcher tested on a three-OS, three-Python CPU matrix plus a self-hosted GPU runner. [Verified] (ecosystem/pickup/_evidence/fine-tuning.md:8
CLI/GUI/data-driven YAML launcher
; ecosystem/pickup/_evidence/fine-tuning.md:233 OS × 3 Python × transformers-compat matrix on CPU
) - axolotl-ai-cloud/axolotl Apache-2.0 AND LicenseRef-Axolotl-Community: a YAML-config launcher for SFT, DPO, LoRA and QLoRA with DeepSpeed and FSDP, CPU tests on every PR and GPU end-to-end tests behind a label. License: Apache-2.0 at the root, but
src/axolotl/integrations/is under the Axolotl Community License, which bars offering fine-tuning to third parties as a service. [Verified] (ecosystem/pickup/_evidence/fine-tuning.md:11YAML-config SFT/DPO/LoRA/QLoRA launcher with DeepSpeed/FSDP
; ecosystem/pickup/_evidence/fine-tuning.md:21GPU e2e in Docker, gated behind run-gpu-tests PR label
; stack/licenses.tsv:7shall not use the Software for any activity that constitutes a commercial activity of offering for free or for sale any services
) - Lightning-AI/litgpt Apache-2.0: hackable pretrain, finetune and eval code whose CPU tests use synthetic micro-configs and download no checkpoint. [Verified] (ecosystem/pickup/_evidence/fine-tuning.md:10
Hackable pretrain/finetune/eval (LoRA, adapters, full)
; ecosystem/pickup/_evidence/fine-tuning.md:40synthetic micro-configs, no checkpoint download at all
)
Copy these practices
- A tiny smoke train on every change: validate the config, load data, train for five steps on a 135M-parameter checkpoint, and assert the checkpoint was written. The pack read one such Axolotl file; the per-family repeats are known by file name only. Starter kit: none. [Verified] (ecosystem/pickup/_evidence/fine-tuning.md:37
full validate_config → load_datasets → train round-trip on HuggingFaceTB/SmolLM2-135M with max_steps: 5
; ecosystem/pickup/_evidence/fine-tuning.md:75did not read every e2e test
) - Exact expected losses at a fixed seed and dtype: commit per-step loss values for a tiny recipe and fail on any deviation; torchtune's version is the evidenced example, from a project no longer maintained. Starter kit: none. [Verified] (ecosystem/pickup/_evidence/fine-tuning.md:46
seed=9, dtype=fp32 overrides; compile=True/False both run
) - Pinned versus latest, side by side: Unsloth runs a pinned small SFT leg next to one on the latest libraries, so an upstream break shows up as a divergence. Starter kit: Incumbent oracle pinned before implementation. [Verified] (ecosystem/pickup/_evidence/fine-tuning.md:60
canary-vs-control legs (tiny Qwen2.5-0.5B SFT pinned vs latest libs)
) - Retry only named transient failures: TRL reruns a failing test up to five times only when its message matches OSError, Timeout, HTTP 502/504 or out of memory; other failures block merges. Starter kit: none. [Verified] (ecosystem/pickup/_evidence/fine-tuning.md:61
--reruns 5 only for OSError/Timeout/HTTP 502/504/OOM messages
) - Config validation as an API, exercised first: the one Axolotl end-to-end test file the pack read calls the config validator and normalizer before training; the pack did not read the others. Starter kit: none. [Verified] (ecosystem/pickup/_evidence/fine-tuning.md:52
calls validate_config(cfg) + normalize_config(cfg) before training
; ecosystem/pickup/_evidence/fine-tuning.md:75did not read every e2e test
) - Evaluate or generate after training: assert the checkpoint actually materializes, then run an eval or generation round-trip. Starter kit: none. [Verified] (ecosystem/pickup/_evidence/fine-tuning.md:67
check_model_output_exists verifies checkpoints actually materialize
) - Distributed logic tested without GPUs: Axolotl runs its DeepSpeed distributed tests on CPU; this covers logic, not numerics. Starter kit: none. [Verified] (ecosystem/pickup/_evidence/fine-tuning.md:58
runs DeepSpeed distributed tests on CPU via
) - A speed win needs the incumbent run side by side: franken_whisper's ledger counts a before/after comparison against yourself as maintenance, and a win only when the incumbent runs in the same invocation with its binary hash recorded; apply it to any launcher speed claim. Starter kit: Incumbent oracle pinned before implementation. [Verified] (packets/franken_whisper-assessment.md:114
requires the actual legacy incumbent running side-by-side in the same invocation, with the incumbent binary's SHA-256 recorded
)
Build only if
- Your deployment cannot carry a Python runtime. The FrankenSuite's Rust ports target that audience but are framework layers at TRL 4, not fine-tuning launchers, so even this constraint points to building on a framework, not to a new launcher. [Inference] (packets/frankenjax-assessment.md:63
Rust developers who need AD without a Python runtime
; synthesis/00-overview.md:103frankentorch | 4 | Explore | Rider
) - The preference and RL correctness gap is a reason to write a differential test harness around an incumbent, not to write a trainer. [Inference] (ecosystem/pickup/pickup-fine-tuning.md:352
No verified conformance oracle for DPO/GRPO correctness beyond
) - License does not justify building. The root-level API read (2026-09-23,
gh api repos/<owner>/<repo> --jq .license.spdx_id) returned Apache-2.0 for all five, but the tree-level census found non-permissive parts: Unsloth's Studio UI and MoE kernels are AGPL-3.0, and Axolotl's integrations bar fine-tuning offered as a service. A fine-tuning-as-a-service product or an AGPL-averse one has a real constraint on those two parts; TRL, LlamaFactory and LitGPT read Apache-2.0 with no conditions noted, so the constraint points to choosing among incumbents, not to a new trainer. [Inference] (stack/licenses.tsv:96unsloth/kernels/moe/LICENSE (MoE grouped-GEMM kernels inside the core package tree) is also AGPL-3.0
; stack/licenses.tsv:7src/axolotl/integrations/ is under the Axolotl Community License
; stack/METHOD.md:79Check the license of everything you tell a builder to adopt
)
Where FrankenSuite touches this
- frankentorch reimplements PyTorch's CPU eager semantics around a Deterministic Autograd Contract (replayable, seeded, provenance-logged gradients) and has an optimizer crate; the companion scopes general training frameworks out of this type, and the packet's reading of it as a reproducible-training base is inference. TRL 4, Explore; license: MIT with a rider barring OpenAI, Anthropic, their affiliates and anyone acting for them. [Verified] (packets/frankentorch-assessment.md:20
organized around the Deterministic Autograd Contract: replayable, seeded, provenance-logged gradient computation
; packets/frankentorch-assessment.md:126optimizers (SGD/Adam/AdamW per README
; packets/frankentorch-assessment.md:299is the reproducible-training workload: an agent team fine-tuning models
; ecosystem/pickup/pickup-fine-tuning.md:8NOT a general ML training framework (PyTorch, JAX)
; packets/frankentorch-assessment.md:102License rider bars OpenAI, Anthropic, affiliates, and anyone acting for them
; synthesis/00-overview.md:103frankentorch | 4 | Explore | Rider
) - frankenjax reimplements JAX's transform semantics and automatic differentiation with a differential conformance harness against JAX fixtures, not the ML framework around them. TRL 4, Explore; license: MIT with a rider barring OpenAI, Anthropic and anyone acting for them. [Verified] (packets/frankenjax-assessment.md:19
not JAX the ML framework, but its mathematical core
; packets/frankenjax-assessment.md:21the license rider bars OpenAI/Anthropic and their agents from even
; synthesis/00-overview.md:90frankenjax | 4 | Explore | Rider
)
What we cannot say
- Whether exact-loss goldens are a category norm: the evidence comes from torchtune, which is sunset (ecosystem/pickup/_evidence/fine-tuning.md:74
torchtune is sunset (2025) — its process evidence (integration markers, golden-loss regression tests)
), and whether they survive FSDP or DeepSpeed is untested (ecosystem/pickup/pickup-fine-tuning.md:348Do exact-loss goldens survive distributed paths (FSDP/DeepSpeed)?
). - How deep Axolotl's testing goes (ecosystem/pickup/_evidence/fine-tuning.md:75
Axolotl process depth unverified in detail
); Unsloth's evidence is mostly installer and compatibility CI (ecosystem/pickup/_evidence/fine-tuning.md:76Unsloth evidence skews toward installer/compat CI
). - Whether the free-GPU notebook legs transfer to a new project (ecosystem/pickup/pickup-fine-tuning.md:359
Whether Kaggle-free-T4 CI legs (unsloth's trick) are available to
). - What tolerance RL training legs need (ecosystem/pickup/pickup-fine-tuning.md:362
Nondeterminism floor for rollout-sampling RL legs (GRPO/PPO)
). - What TRL's
tests/invariantdirectory asserts was not read, so a partial correctness oracle may exist there (ecosystem/pickup/_evidence/fine-tuning.md:30tests/ has matching dir layout (tests/distributed, tests/invariant, per-trainer files)
). - Axolotl's organisation moved and LlamaFactory was renamed; the pointers use the live paths (ecosystem/pickup/_evidence/fine-tuning.md:79
LLaMA-Factory is now hiyouga/LlamaFactory (case-renamed)
). - OpenRLHF and LMOps were checked for activity only (ecosystem/pickup/_evidence/fine-tuning.md:77
OpenRLHF / LMOps are trend-only
). Star counts are a snapshot (ecosystem/pickup/_evidence/fine-tuning.md:78Star counts are snapshot
). Workflow files were read, not observed running; licenses come from the author's fresh read, not the pack.
Revisit when
- A published conformance suite for DPO, GRPO or PPO trainers appears, or TRL adopts exact-loss regression tests.
- An incumbent launcher loses maintenance the way torchtune did.
- frankentorch ships a release and gains an outside consumer that fine-tunes on it.
Sources
- Pickup companion: ecosystem/pickup/pickup-fine-tuning.md, the charter seed, oracle candidates, and first claim registry for this area.
- Evidence pack: ecosystem/pickup/_evidence/fine-tuning.md, the repositories checked against the GitHub API, with dates.
- Verdict source: stack/fine-tuning.md, written under stack/METHOD.md.
Starting a build here?
Begin from the pickup companion for Fine-tuning infrastructure: it carries the charter seed, the oracle candidates and how to verify them, and the first claims to register, so a new project starts from evidence instead of from scratch.