Metadata-Version: 2.5
Name: ftdoctor
Version: 0.1.0
Summary: Did your fine-tune overfit? Which checkpoint should you keep? What can you delete? Reads any HuggingFace Trainer run, local or on the Hub.
Project-URL: Homepage, https://github.com/junglezke/ftdoctor
Project-URL: Repository, https://github.com/junglezke/ftdoctor
Project-URL: Issues, https://github.com/junglezke/ftdoctor/issues
Author: junglezke
License: 
                                         Apache License
                                   Version 2.0, January 2004
                                http://www.apache.org/licenses/
        
           TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
        
           1. Definitions.
        
              "License" shall mean the terms and conditions for use, reproduction,
              and distribution as defined by Sections 1 through 9 of this document.
        
              "Licensor" shall mean the copyright owner or entity authorized by
              the copyright owner that is granting the License.
        
              "Legal Entity" shall mean the union of the acting entity and all
              other entities that control, are controlled by, or are under common
              control with that entity. For the purposes of this definition,
              "control" means (i) the power, direct or indirect, to cause the
              direction or management of such entity, whether by contract or
              otherwise, or (ii) ownership of fifty percent (50%) or more of the
              outstanding shares, or (iii) beneficial ownership of such entity.
        
              "You" (or "Your") shall mean an individual or Legal Entity
              exercising permissions granted by this License.
        
              "Source" form shall mean the preferred form for making modifications,
              including but not limited to software source code, documentation
              source, and configuration files.
        
              "Object" form shall mean any form resulting from mechanical
              transformation or translation of a Source form, including but
              not limited to compiled object code, generated documentation,
              and conversions to other media types.
        
              "Work" shall mean the work of authorship, whether in Source or
              Object form, made available under the License, as indicated by a
              copyright notice that is included in or attached to the work
              (an example is provided in the Appendix below).
        
              "Derivative Works" shall mean any work, whether in Source or Object
              form, that is based on (or derived from) the Work and for which the
              editorial revisions, annotations, elaborations, or other modifications
              represent, as a whole, an original work of authorship. For the purposes
              of this License, Derivative Works shall not include works that remain
              separable from, or merely link (or bind by name) to the interfaces of,
              the Work and Derivative Works thereof.
        
              "Contribution" shall mean any work of authorship, including
              the original version of the Work and any modifications or additions
              to that Work or Derivative Works thereof, that is intentionally
              submitted to Licensor for inclusion in the Work by the copyright owner
              or by an individual or Legal Entity authorized to submit on behalf of
              the copyright owner. For the purposes of this definition, "submitted"
              means any form of electronic, verbal, or written communication sent
              to the Licensor or its representatives, including but not limited to
              communication on electronic mailing lists, source code control systems,
              and issue tracking systems that are managed by, or on behalf of, the
              Licensor for the purpose of discussing and improving the Work, but
              excluding communication that is conspicuously marked or otherwise
              designated in writing by the copyright owner as "Not a Contribution."
        
              "Contributor" shall mean Licensor and any individual or Legal Entity
              on behalf of whom a Contribution has been received by Licensor and
              subsequently incorporated within the Work.
        
           2. Grant of Copyright License. Subject to the terms and conditions of
              this License, each Contributor hereby grants to You a perpetual,
              worldwide, non-exclusive, no-charge, royalty-free, irrevocable
              copyright license to reproduce, prepare Derivative Works of,
              publicly display, publicly perform, sublicense, and distribute the
              Work and such Derivative Works in Source or Object form.
        
           3. Grant of Patent License. Subject to the terms and conditions of
              this License, each Contributor hereby grants to You a perpetual,
              worldwide, non-exclusive, no-charge, royalty-free, irrevocable
              (except as stated in this section) patent license to make, have made,
              use, offer to sell, sell, import, and otherwise transfer the Work,
              where such license applies only to those patent claims licensable
              by such Contributor that are necessarily infringed by their
              Contribution(s) alone or by combination of their Contribution(s)
              with the Work to which such Contribution(s) was submitted. If You
              institute patent litigation against any entity (including a
              cross-claim or counterclaim in a lawsuit) alleging that the Work
              or a Contribution incorporated within the Work constitutes direct
              or contributory patent infringement, then any patent licenses
              granted to You under this License for that Work shall terminate
              as of the date such litigation is filed.
        
           4. Redistribution. You may reproduce and distribute copies of the
              Work or Derivative Works thereof in any medium, with or without
              modifications, and in Source or Object form, provided that You
              meet the following conditions:
        
              (a) You must give any other recipients of the Work or
                  Derivative Works a copy of this License; and
        
              (b) You must cause any modified files to carry prominent notices
                  stating that You changed the files; and
        
              (c) You must retain, in the Source form of any Derivative Works
                  that You distribute, all copyright, patent, trademark, and
                  attribution notices from the Source form of the Work,
                  excluding those notices that do not pertain to any part of
                  the Derivative Works; and
        
              (d) If the Work includes a "NOTICE" text file as part of its
                  distribution, then any Derivative Works that You distribute must
                  include a readable copy of the attribution notices contained
                  within such NOTICE file, excluding those notices that do not
                  pertain to any part of the Derivative Works, in at least one
                  of the following places: within a NOTICE text file distributed
                  as part of the Derivative Works; within the Source form or
                  documentation, if provided along with the Derivative Works; or,
                  within a display generated by the Derivative Works, if and
                  wherever such third-party notices normally appear. The contents
                  of the NOTICE file are for informational purposes only and
                  do not modify the License. You may add Your own attribution
                  notices within Derivative Works that You distribute, alongside
                  or as an addendum to the NOTICE text from the Work, provided
                  that such additional attribution notices cannot be construed
                  as modifying the License.
        
              You may add Your own copyright statement to Your modifications and
              may provide additional or different license terms and conditions
              for use, reproduction, or distribution of Your modifications, or
              for any such Derivative Works as a whole, provided Your use,
              reproduction, and distribution of the Work otherwise complies with
              the conditions stated in this License.
        
           5. Submission of Contributions. Unless You explicitly state otherwise,
              any Contribution intentionally submitted for inclusion in the Work
              by You to the Licensor shall be under the terms and conditions of
              this License, without any additional terms or conditions.
              Notwithstanding the above, nothing herein shall supersede or modify
              the terms of any separate license agreement you may have executed
              with Licensor regarding such Contributions.
        
           6. Trademarks. This License does not grant permission to use the trade
              names, trademarks, service marks, or product names of the Licensor,
              except as required for reasonable and customary use in describing the
              origin of the Work and reproducing the content of the NOTICE file.
        
           7. Disclaimer of Warranty. Unless required by applicable law or
              agreed to in writing, Licensor provides the Work (and each
              Contributor provides its Contributions) on an "AS IS" BASIS,
              WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
              implied, including, without limitation, any warranties or conditions
              of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
              PARTICULAR PURPOSE. You are solely responsible for determining the
              appropriateness of using or redistributing the Work and assume any
              risks associated with Your exercise of permissions under this License.
        
           8. Limitation of Liability. In no event and under no legal theory,
              whether in tort (including negligence), contract, or otherwise,
              unless required by applicable law (such as deliberate and grossly
              negligent acts) or agreed to in writing, shall any Contributor be
              liable to You for damages, including any direct, indirect, special,
              incidental, or consequential damages of any character arising as a
              result of this License or out of the use or inability to use the
              Work (including but not limited to damages for loss of goodwill,
              work stoppage, computer failure or malfunction, or any and all
              other commercial damages or losses), even if such Contributor
              has been advised of the possibility of such damages.
        
           9. Accepting Warranty or Additional Liability. While redistributing
              the Work or Derivative Works thereof, You may choose to offer,
              and charge a fee for, acceptance of support, warranty, indemnity,
              or other liability obligations and/or rights consistent with this
              License. However, in accepting such obligations, You may act only
              on Your own behalf and on Your sole responsibility, not on behalf
              of any other Contributor, and only if You agree to indemnify,
              defend, and hold each Contributor harmless for any liability
              incurred by, or claims asserted against, such Contributor by reason
              of your accepting any such warranty or additional liability.
        
           END OF TERMS AND CONDITIONS
        
           APPENDIX: How to apply the Apache License to your work.
        
              To apply the Apache License to your work, attach the following
              boilerplate notice, with the fields enclosed by brackets "[]"
              replaced with your own identifying information. (Don't include
              the brackets!)  The text should be enclosed in the appropriate
              comment syntax for the file format. We also recommend that a
              file or class name and description of purpose be included on the
              same "printed page" as the copyright notice for easier
              identification within third-party archives.
        
           Copyright [yyyy] [name of copyright owner]
        
           Licensed under the Apache License, Version 2.0 (the "License");
           you may not use this file except in compliance with the License.
           You may obtain a copy of the License at
        
               http://www.apache.org/licenses/LICENSE-2.0
        
           Unless required by applicable law or agreed to in writing, software
           distributed under the License is distributed on an "AS IS" BASIS,
           WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
           See the License for the specific language governing permissions and
           limitations under the License.
License-File: LICENSE
Keywords: checkpoint,early-stopping,fine-tuning,huggingface,llm,lora,mlops,overfitting,qlora,sft,trainer,transformers,trl,unsloth
Classifier: Development Status :: 4 - Beta
Classifier: Intended Audience :: Developers
Classifier: Intended Audience :: Science/Research
Classifier: License :: OSI Approved :: Apache Software License
Classifier: Programming Language :: Python :: 3
Classifier: Programming Language :: Python :: 3.10
Classifier: Programming Language :: Python :: 3.11
Classifier: Programming Language :: Python :: 3.12
Classifier: Programming Language :: Python :: 3.13
Classifier: Programming Language :: Python :: 3.14
Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
Requires-Python: >=3.10
Requires-Dist: numpy>=1.21
Provides-Extra: dev
Requires-Dist: pytest-cov>=4.0; extra == 'dev'
Requires-Dist: pytest>=7.0; extra == 'dev'
Requires-Dist: ruff>=0.6; extra == 'dev'
Description-Content-Type: text/markdown

<div align="center">

# ftdoctor

**Did your fine-tune overfit? Which checkpoint should you keep? What can you delete?**

Point it at a training output folder — or at any model on the HuggingFace Hub — and get
the answer in a second.

[![CI](https://github.com/junglezke/ftdoctor/actions/workflows/ci.yml/badge.svg)](https://github.com/junglezke/ftdoctor/actions/workflows/ci.yml)
[![PyPI](https://img.shields.io/pypi/v/ftdoctor.svg)](https://pypi.org/project/ftdoctor/)
[![Python](https://img.shields.io/pypi/pyversions/ftdoctor.svg)](https://pypi.org/project/ftdoctor/)
[![License](https://img.shields.io/badge/license-Apache%202.0-blue.svg)](LICENSE)

<img src="docs/assets/report.svg" alt="ftdoctor on a public Mistral-7B fine-tune: keep checkpoint-500; the best eval at step 900 was never saved; the last eval is 6% worse" width="100%">

</div>

---

Works with anything built on the HuggingFace `Trainer` — **TRL's `SFTTrainer`, Unsloth,
LLaMA-Factory, Axolotl**, `Seq2SeqTrainer` for Whisper, plain `Trainer` for BERT. They all
leave the full training history in `trainer_state.json`, and almost nobody reads it.

## Try it on a public model — no install of anything else, no GPU

```bash
pip install ftdoctor
ftdoctor anshpunia8597/mistral-7b-medical-qa-finetuned
```

Or the latest from `main`: `pip install git+https://github.com/junglezke/ftdoctor`.

That is the banner above: a real Mistral-7B fine-tune on the Hub. ftdoctor reads its
`trainer_state.json` and the checkpoint folders the repo actually holds, and says:

- the best eval was at **step 900 — which was never saved**, so you cannot have it;
- of the checkpoints that exist, **`checkpoint-500`** is the best;
- the last eval is **6% worse**, and that is outside the eval's own noise, so it is real.

## On your own run

```bash
ftdoctor ./outputs                    # the folder with checkpoint-*/ in it
ftdoctor ./outputs/checkpoint-1200    # or any one checkpoint
ftdoctor trainer_state.json           # or just the file
```

```python
from ftdoctor import load, diagnose

result = diagnose(load("./outputs"))
print(result.headline)          # outputs: keep checkpoint-1000; overfitting.
print(result.pick.keep_step)    # 1000
```

## And get your disk back

A full fine-tune that saves every few hundred steps leaves dozens of multi-gigabyte
checkpoints behind. Most of each one is optimizer state you only need to *resume*
training.

```bash
ftdoctor prune ./outputs            # show the plan -- deletes nothing
ftdoctor prune ./outputs --apply    # do it
```

By default `prune` keeps the recommended checkpoint and the latest one untouched, and from
every other checkpoint removes only resume-only files — `optimizer.pt`, `scheduler.pt`,
RNG state, DeepSpeed shards. **Every model stays loadable.** `--mode full` deletes the other
checkpoints outright, and is refused when the run has no eval to base a recommendation on.

Before deleting anything, `--apply` re-checks each target is a `checkpoint-N` directory
directly inside your output folder, containing a `trainer_state.json`, and not one being
kept.

## What 89 public fine-tunes look like

To calibrate the checks, ftdoctor was run on **89 `trainer_state.json` files** published on
the Hub — LLM SFT and QLoRA runs, Whisper, BERT, audio classifiers. 88 were supervised
fine-tunes; one turned out to be an RL run (more on that below).

| | |
|---|---|
| **41 of 88 (47%)** | logged **no eval metric at all**. Nothing can tell whether they overfit — training loss only ever goes down. |
| **23** | had fewer than 5 evals: too few to tell a trend from noise. |
| **24** | had enough evals to judge. Of those: |
| &nbsp;&nbsp;10 | the final checkpoint was the best one. |
| &nbsp;&nbsp;**10** | the final checkpoint was *not* the minimum — **but the difference was within the eval's own noise**. A naive `argmin(eval_loss)` tells these people to go back to an earlier checkpoint for nothing. |
| &nbsp;&nbsp;4 | the final checkpoint was genuinely worse; 2 of them by more than 3%. |

And a few things only real logs turn up:

- **Loss that doubles at every epoch boundary.** A Qwen2.5-7B SFT whose loss jumped on the
  first step of each of its 10 epochs. Periodic, so not instability — the step straddling
  the epoch boundary computes its loss differently. The usual advice for loss spikes (lower
  the learning rate) would have been wrong.
- **Loss of exactly 0.0 from the first step — which was fine.** A Gemma-3-12B run whose
  name said SFT logged a loss of 0.0 from step one. An early version of ftdoctor called it
  "trained on nothing, labels masked". It was a GRPO run, where a loss near zero is normal.
  ftdoctor now recognises RL runs, skips the loss-based checks instead of misreading them,
  and hands them to [rldoctor](https://github.com/junglezke/rldoctor). (For a genuinely
  zero loss in SFT, the usual cause *is* every label masked to `-100` — and ftdoctor says so.)
- **Eval loss up 37%, accuracy still improving.** An audio classifier selected on accuracy.
  Judged on loss it overfit badly; judged on the metric its author actually chose, the last
  checkpoint was the best. ftdoctor recovers that choice from `best_metric` and says so.
- **`inf` gradients that do not matter.** A Mistral run logged `grad_norm = inf` at 10
  scattered steps out of 34,000. Under fp16 that is the loss scaler skipping an overflowing
  step — harmless, and telling someone to roll back over it would be wrong.

## What it checks

| | catches |
|---|---|
| **Which checkpoint** | best eval, the noise band around it, and the best checkpoint *that exists* |
| **Overfitting** | eval getting worse while training loss keeps falling; the train/eval gap; steps and hours spent after the best point |
| **Stopped improving** | the eval flatlined early: same result for a fraction of the compute |
| **Loss vs your metric** | eval loss rising while accuracy / WER / BLEU improves — not overfitting in the sense you care about |
| **No eval** | and the exact `TrainingArguments` to fix it |
| **Broken loss** | NaN loss; loss of exactly zero (masked labels) |
| **RL runs** | recognised and handed to rldoctor, not misread as a broken SFT |
| **Not learning** | loss flat or rising, with the learning rate that explains it |
| **Loss spikes** | real spikes, periodic ones, and ones that land on epoch boundaries — each with different advice |
| **Gradients** | exactly-zero gradients (nothing trainable), spikes that the loss felt, fp16 overflow that it didn't |
| **Learning-rate schedule** | a tail of steps trained at a learning rate of ~0 |
| **Eval below train** | eval loss consistently under training loss — check for train/eval overlap |
| **Unfinished runs** | the log ends before `max_steps` |

Every finding comes with the numbers it fired on and a concrete change to make.

## How it decides

**Noise-aware, not `argmin`.** Each eval metric has its own step-to-step noise, estimated
robustly from successive differences. Any checkpoint within 2× that noise (or 1%, whichever
is larger) of the best is reported as equivalent. "Your last checkpoint is 1.2% worse" is
only said when 1.2% is more than the eval's jitter.

**Your metric, not ours.** `trainer_state.json` records `best_metric` but not which metric
it was. ftdoctor finds the eval series that contains that exact value at `best_global_step`
and ranks on it — accuracy, WER, BLEU, whatever you selected on. Direction (higher or lower
is better) comes from the name, and is checked against which extreme the Trainer kept. Use
`--metric` to override.

**Only checkpoints that exist.** The best eval often lands on a step that was never saved,
or was deleted by `save_total_limit`. Checkpoints are taken from the folders really present
— on disk or on the Hub — never inferred from `save_steps`, which is meaningless under
`save_strategy="epoch"`. Given only a bare `trainer_state.json`, ftdoctor says it cannot
tell which checkpoints exist instead of guessing.

**Calibrated on real runs.** Every threshold was set against the 89 public runs above, and
several were changed because of them: a 1.5× loss-spike rule flagged ordinary small-batch
noise; a band-only overfitting rule called 1% drops "overfitting"; a global-median gradient
rule mistook the normal early-training decline for spikes. Ten of those runs are pinned in
[`validation/`](validation/runs.json) and re-checked in CI against their public sources.

## What it will not do

- **It only sees what was logged.** No eval metric, no overfitting verdict — it will tell
  you so rather than guess from training loss.
- **It does not run your model.** It reads the history. Whether the "best" checkpoint is good
  enough for your task is still a question for your own evaluation.
- **It cannot see sampling noise it was not shown.** The noise band comes from the eval's
  step-to-step variation. A tiny eval set can be noisy in ways a smooth curve hides; if yours
  has a few dozen examples, treat small differences with suspicion either way.
- **`prune` cannot be undone.** That is why it defaults to a dry run, defaults to keeping
  every model's weights, and checks every path before touching it.

## In CI

```bash
ftdoctor ./outputs --fail-on warning --format markdown -o "$GITHUB_STEP_SUMMARY"
```

Exits non-zero when a finding at that severity fires, and writes a Markdown report for the
job summary.

## Related

ftdoctor is for supervised fine-tuning. For **RL post-training** (GRPO, PPO, RLVR) — reward
hacking, entropy collapse, degenerate groups — see
[**rldoctor**](https://github.com/junglezke/rldoctor), and to test the reward function
itself, [**rewardlint**](https://github.com/junglezke/rewardlint).

## Development

```bash
git clone https://github.com/junglezke/ftdoctor && cd ftdoctor
pip install -e ".[dev]"
pytest                                                      # 59 tests
python validation/fetch.py && python validation/run.py --check   # the public runs
python tools/make_banner.py                                 # regenerate the README image
```

## Contributing

**The most useful thing you can send is a run where ftdoctor is wrong** — it called a fine
run broken, or missed something real. A link to a public model with a `trainer_state.json`
is enough. There is an issue template for it.

## License

Apache-2.0.
