Skip to content

Better data. Better model.

dropoutt scan ./data  done Scanned 332 records   ◧◨ 5 problems would fail a fine-tuning run   332 records   ·   5 datasets   ·   15k tokens   ·   tr 97%, unknown 3%   What would go wrong     would block  Role names your trainer will not recognise  T0-ROLE-002      60 of 291 records · 21%      non-canonical role names: 'human' (60), 'gpt' (60)      → Map roles to system/user/assistant before training, or confirm yourtrainer maps them.     would block  Files are not training data  T0-SCHEMA-001      1 of 5 datasets      logs: records carry type='queue-operation'      → Point the scan at your dataset directory rather than a project or logdirectory.     warning  Responses that teach the model nothing  T0-DEGEN-001      133 of 324 responses · 41%      2 trivial, 131 looping responses      → Remove trivial, looping or prompt-copying responses.   Written to data/.dropoutt    report.html       the excerpts and every finding in full    findings.jsonl    one finding per line, for scripts    fingerprint.json  comparable measurements, for diffing runs  No pass-or-fail verdict: no target declared. Pass --target sft to turnfindings into an exit code.  332 records in 0.4s.

Audit your LLM datasets for blank loss masks, benchmark leaks, PII, and repetition loops. Run it locally in seconds before you spend on GPU compute.

pip install dropoutt
Read the docs

See what your data covers.

A model can only learn the topics its data contains. Atlas shows where your dataset goes deep, where it runs thin, and what it leaves out, so you can fill the gaps before you pay for a training run.

dropoutt atlas ./data  done Placed 2,090 of 2,100 records   ◧◨ Where this corpus sits on the map     atlas-v3   2,100 records   ·   7 datasets   ·   en 85%, tr 14%, fr under 0.1%     Effective coverage 1004.4 of 4,096 (1,005 subregions hold any records)(broad)    2,090 placed of 2,100 sampled records · 10 off the map (0.5%)     Machine learning, statistics and computational methods  19/19 reach · 11.1%share · 231 records         20×  Algorithms, optimiza…          3.4×  Mathematics, analysi…         12×  Probability, estimat…          3.2×  Dynamical systems, c…     Mathematics, algorithms and programming problems  16/20 reach · 8.5% share ·177 records         33×  Python, NumPy and da…          2.5×  Graph theory, graphs…         14×  Code and forum posts…          2.1×  Mathematical functio…     Programming, web code and desktop software how-tos  17/20 reach · 4.5% share· 94 records         14×  Programming, scripts…          1.7×  Plugins, WordPress t…        3.8×  HTML, CSS and front-…          1.5×  Software projects, b…     223 further areas reached; their subregions are named in the report files    25 of the map's 256 subject areas never reached     Of the 1005 places you reach, 864 hold 50% of your data between them      Real presence in 141 places, a toehold in the rest. An occupancy countreads a place holding one record the same as one holding a third of the corpus,which is how a narrow corpus comes to look broad.   Written to corpus/.dropoutt    atlas.html  the map, drawn  2,100 records in 1.4s. Run `dropoutt scan` for the checks.

256 topics. 4,096 subtopics.

no reachas common as on the mapover-represented, up to 33×

Machine learning, statistics and computational methods

231 records11.1% of placed

19 / 19 reach
Subtopicvs the map
Algorithms, optimization and computational complexity20×, 20× the map's own density
Probability, estimation and stochastic processes12×, 12× the map's own density
Machine learning, reinforcement learning and training algorithms10×, 10× the map's own density
Optimization, sparse methods and numerical algorithms10×, 10× the map's own density
Hypothesis testing, regression and applied statistics6.9×, 6.9× the map's own density
Deep neural networks and their applications5.7×, 5.7× the map's own density
Graphs, networks and graph neural methods5.0×, 5.0× the map's own density

12 more subtopics

Hotels, lodging and guest reviews

86 records4.1% of placed

14 / 22 reach
Subtopicvs the map
TripAdvisor hotel reviews and booking-partner pages19×, 19× the map's own density
City, airport and chain hotels4.0×, 4.0× the map's own density
Hotels, rooms and guest-house listings3.0×, 3.0× the map's own density
Beach resorts, seaside hotels and oceanfront stays2.4×, 2.4× the map's own density
Hotel booking, deals and cheap-rate search pages2.3×, 2.3× the map's own density
Hotel listings of rooms, location and facilities2.0×, 2.0× the map's own density
Hotel rates, reservations and stay conditions2.0×, 2.0× the map's own density

15 more subtopics

Baking, desserts and sweet recipes

37 records1.8% of placed

17 / 22 reach
Subtopicvs the map
Cookies and cookie baking3.8×, 3.8× the map's own density
Ice cream, gelato and frozen desserts3.0×, 3.0× the map's own density
Measured baking recipes and dessert formulas2.7×, 2.7× the map's own density
Fruit desserts, pumpkin baking and autumn sweets2.0×, 2.0× the map's own density
Tea, matcha and tea brewing1.8×, 1.8× the map's own density
Bread, buns and yeast baking1.6×, 1.6× the map's own density
Cakes, cupcakes and cake decorating1.6×, 1.6× the map's own density

15 more subtopics

Crafts, printing and handmade goods

58 records2.8% of placed

12 / 17 reach
Subtopicvs the map
Printers, 3D printing and print production18×, 18× the map's own density
Rubber-stamped cards and papercraft stamping1.9×, 1.9× the map's own density
Greeting cards and stationery1.7×, 1.7× the map's own density
Paper, felt and small handmade craft projects1.6×, 1.6× the map's own density
Children's coloring, crafts and art activities1.3×, 1.3× the map's own density
Drawing, sketching and drawing tools1.3×, 1.3× the map's own density
Graphic design, logos and design templates1.2×, 1.2× the map's own density

10 more subtopics

Browse every topic

Find the records that break training.

Point dropoutt scan at a dataset folder and 34 checks run at once. Each finding lists the records behind it and a suggested fix.

  • Empty loss masks

    Records that add zero trainable tokens. You pay to process them and the model learns nothing.

  • Duplicates

    Exact copies, near copies, and the same prompt answered two different ways.

  • Benchmark leaks

    Training records that overlap evaluation sets and inflate the scores you report.

  • Truncated records

    Records longer than your sequence length, and answers cut off at a generation cap.

  • Personal data and secrets

    Emails, phone numbers, and credentials in training text. Every report masks the match.

  • Repetition loops

    Looping, trivial, or prompt-copying responses that a model learns to repeat.

See all 36 checks

Works on your laptop and in CI.

One install, no account, no server to run.

  • Private by default

    Your data stays on your machine. Fetch the models once, then scan with the network switched off.

  • Reports you can share

    An HTML report with excerpts and a suggested fix for each finding, plus JSON for your own scripts.

  • A gate in CI

    Declare what you are training and a blocking finding fails the build with exit code 10.

Run your first scan.

pip install dropoutt
uv tool install dropoutt

Free and open source under Apache 2.0 · Python 3.10 or newer