File size: 26,965 Bytes
52de1e3 521e47c 52de1e3 1017e80 98c2631 1017e80 84e0f39 98c2631 84e0f39 521e47c 1017e80 98c2631 1017e80 28fbbeb 84e0f39 28fbbeb 84e0f39 28fbbeb 98c2631 1017e80 84e0f39 1017e80 84e0f39 1017e80 98c2631 84e0f39 1017e80 84e0f39 98c2631 84e0f39 98c2631 1017e80 84e0f39 98c2631 84e0f39 98c2631 84e0f39 98c2631 84e0f39 1017e80 84e0f39 1017e80 84e0f39 b6859cb 98c2631 1017e80 98c2631 1017e80 84e0f39 1017e80 84e0f39 1017e80 84e0f39 98c2631 84e0f39 1017e80 28fbbeb 84e0f39 98c2631 28fbbeb 98c2631 28fbbeb 98c2631 28fbbeb 98c2631 28fbbeb 84e0f39 98c2631 84e0f39 521e47c 98c2631 52de1e3 98c2631 52de1e3 98c2631 52de1e3 98c2631 52de1e3 98c2631 5c4f2fd 52de1e3 0bcd4fa 98c2631 52de1e3 98c2631 52de1e3 98c2631 5c4f2fd 52de1e3 0bcd4fa 1e871c4 52de1e3 011bc8a 98c2631 011bc8a 52de1e3 98c2631 52de1e3 98c2631 52de1e3 98c2631 52de1e3 0bcd4fa 5c4f2fd 98c2631 52de1e3 98c2631 0bcd4fa | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 | ---
viewer: false
tags: [uv-script, classification, fine-tuning, few-shot, zero-shot, decision-model, setfit, gliner2, kev, vllm, structured-outputs, hf-jobs]
---
# Classification Scripts
Text classification on [HF Jobs](https://huggingface.co/docs/hub/jobs): label a dataset with a model that needs no training, or train your own classifier from labelled examples.
If you have seen [Jev](https://docs.typesafe.ai/introduction) and other "System One" models: the
models these scripts train are small, open versions of the same idea. They read a piece of data and
return a label with a probability, and you can train one on your own labels. For example, this
[demo](https://huggingface.co/spaces/davanstrien/hub-task-tagger) suggests task tags for any Hub
dataset; its model was fine-tuned with `train-gliner2.py` in 17 minutes.
To try it on your own account, this command fine-tunes a classifier for British Library book titles
(Fiction / Non-fiction). Training takes about 2 minutes on a `t4-small` and costs about $0.02.
Accuracy goes from 0.767 zero-shot to 0.907 fine-tuned. You get a private model repo, and its card
shows both scores next to the majority-class baseline. Copy and paste it as it is: the model goes
to your own account.
```bash
hf jobs uv run --flavor t4-small --timeout 1h --secrets HF_TOKEN \
https://huggingface.co/datasets/uv-scripts/classification/raw/main/train-gliner2.py \
biglam/blbooksgenre gliner2-blbooks-genre \
--dataset-config title_genre_classifiction --text-column title
```
You need the `hf` CLI, signed in, and Jobs credit: see the [Jobs quickstart](https://huggingface.co/docs/hub/jobs-quickstart).
| Script | What it does |
|--------|--------------|
| [`classify-gliner2.py`](#zero-shot-first-then-fine-tune-gliner2) | **Label a dataset** with GLiNER2: zero-shot from label names, or with a `train-gliner2.py` model |
| [`train-gliner2.py`](#zero-shot-first-then-fine-tune-gliner2) | **Fine-tune** [GLiNER2](https://huggingface.co/fastino), a small model (74M–287M) that already classifies zero-shot, and report the zero-shot score next to the fine-tuned one |
| [`train-setfit.py`](#few-shot-with-setfit-train-setfitpy) | **Few-shot** train a classifier from 8-64 labels per class with [SetFit](https://huggingface.co/docs/setfit) — runs on CPU or GPU |
| [`train-classifier.py`](#fine-tune-a-classifier-train-classifierpy) | **Fine-tune** an encoder into a classifier (default: [LFM2.5-Encoder-350M](https://huggingface.co/LiquidAI/LFM2.5-Encoder-350M)) and push it to the Hub |
| [`train-kev.py`](#fine-tune-a-kev-decision-model-train-kevpy) | **Fine-tune** the open [Kev-4B](https://huggingface.co/jaredpalmer/kev-4b) decision model on your own questions, with Kev's own training code |
| [`classify-dataset.py`](#zero-shot-classification-classify-datasetpy) | **Zero-shot** classify a dataset with an instruction LLM (SmolLM3 + vLLM, structured outputs) |
| [`classify-dataset-sglang.py`](#zero-shot-classification-classify-datasetpy) | Zero-shot variant on SGLang (reasoning-aware `<think>` models) |
Pick by how many labels you have:
| Labels you have | Use | Hardware |
|---|---|---|
| none | `classify-gliner2.py --labels ...` for a cheap first pass; `classify-dataset.py` when the task needs an LLM's reasoning | small GPU (CPU works at ~1.4 rows/s); GPU |
| ~8-64 per class | `train-setfit.py` | CPU supported; GPU for faster training |
| a few hundred to a few thousand | `train-gliner2.py`, which also shows you what zero-shot already gets | small GPU (`t4-small`) |
| a few thousand or more | `train-classifier.py` | GPU |
The rungs chain: bootstrap labels with `classify-gliner2.py` or `classify-dataset.py`, review them,
then train a small dedicated model on what you kept. Each rung is one command, so you can move up
as you collect more labels.
## Zero-shot first, then fine-tune (GLiNER2)
Label a dataset with your own list of labels, see how far zero-shot gets you, then fine-tune a
small model on your labels, in minutes and for a few cents on one GPU. The result is a model
that returns a label and a probability for every row, and is small enough to run on a CPU.
[GLiNER2](https://github.com/fastino-ai/GLiNER2) ([models from Fastino](https://huggingface.co/fastino)) is a small encoder that reads the label names
as part of its input, so it classifies with no training at all, and fine-tuning teaches it what
your labels mean in your data. (For entity extraction with the original GLiNER library, see
[`uv-scripts/gliner`](https://huggingface.co/datasets/uv-scripts/gliner).) Two scripts:
- **`train-gliner2.py`** scores the base model zero-shot, fine-tunes it on your labels, and
scores it again on the same held-out rows. The model card reports both scores next to the
majority-class baseline.
- **`classify-gliner2.py`** labels a whole dataset. Pass `--labels` for zero-shot, or `--model`
with a `train-gliner2.py` output; the tasks and labels are read from the model repo.
Labels are passed as separate words; quote a label with spaces:
`--labels World Sports Business "Science and technology"`. A fine-tuned model is passed
by name: `--model gliner2-blbooks-genre`. A name on its own means your account; use `org/name`
for an organisation.
### Quick start
```bash
# fine-tune: British Library book titles -> Fiction / Non-fiction
hf jobs uv run --flavor t4-small --timeout 1h --secrets HF_TOKEN \
https://huggingface.co/datasets/uv-scripts/classification/raw/main/train-gliner2.py \
biglam/blbooksgenre gliner2-blbooks-genre \
--dataset-config title_genre_classifiction --text-column title
# label a dataset with that model
hf jobs uv run --flavor t4-small --timeout 1h --secrets HF_TOKEN \
https://huggingface.co/datasets/uv-scripts/classification/raw/main/classify-gliner2.py \
biglam/blbooksgenre blbooks-genre-predictions \
--dataset-config title_genre_classifiction --text-column title --model gliner2-blbooks-genre
# or skip training: zero-shot from label names
hf jobs uv run --flavor t4-small --timeout 1h --secrets HF_TOKEN \
https://huggingface.co/datasets/uv-scripts/classification/raw/main/classify-gliner2.py \
fancyzhx/ag_news ag-news-topics --split test \
--labels World Sports Business "Science and technology" --task-name topic
```
The second command labels the same rows the model was trained on, so it shows the workflow, not
the model's accuracy; the held-out scores are on the model card. For real use, point it at data
the model has not seen. Outputs are **private by default** (`--public` to opt out).
These commands use the default base model, `fastino/gliner2.5-multi-v1` (multilingual). For
English text, `--base-model fastino/gliner2.5-base-v1` is smaller and faster;
`gliner2.5-small-v1` is the fastest and loses about 4 points on the 52-tag example. For English zero-shot, try
`fastino/GLiNER2.5-Decide`. Sizes and speeds: [Choosing a model size](GLINER2-NOTES.md#choosing-a-model-size).
### Results
Fine-tuning beat zero-shot on every dataset below. The public-dataset runs took 2 to 30 minutes of
training on one GPU and cost $0.01 to $0.20; the 52-tag example took 17 minutes and about $1.50.
| Dataset | Task | Labels | Train rows × epochs | Train time | Train cost | Metric | Majority floor | Zero-shot | Fine-tuned |
|---|---|---|---|---|---|---|---|---|---|
| [`biglam/blbooksgenre`](https://huggingface.co/datasets/biglam/blbooksgenre) (book titles) | single-label | 2 | 1,562 × 5 | 141s | $0.02 | accuracy | 0.747 | 0.767 (0.753–0.782) | **0.907** (0.897–0.925) |
| [`fancyzhx/ag_news`](https://huggingface.co/datasets/fancyzhx/ag_news) | single-label | 4 | 2,000 × 2 | 125s | $0.01 | accuracy | 0.268 | 0.718 | **0.852** |
| [`google-research-datasets/go_emotions`](https://huggingface.co/datasets/google-research-datasets/go_emotions) | multi-label | 28 | 2,000 × 2 | 216s | $0.02 | micro F1 | — | 0.265 | **0.464** |
| [`SetFit/TREC-QC`](https://huggingface.co/datasets/SetFit/TREC-QC), two tasks in one model | single-label ×2 | 6 + 50 | 5,452 × 3 | 1,761s | $0.20 | accuracy | 0.276 / 0.246 | 0.542 / 0.468 | **0.954 / 0.876** |
| same, on `a10g-small`, default batch size, bf16 | single-label ×2 | 6 + 50 | 5,452 × 3 | 509s | $0.14 | accuracy | 0.276 / 0.246 | not run | **0.944 / 0.872** |
| [`stanfordnlp/imdb`](https://huggingface.co/datasets/stanfordnlp/imdb) (reviews; 15% truncated at 2,000 characters) | single-label | 2 | 1,000 × 1 | 122s | $0.01 | accuracy | 0.500 | 0.777 | **0.840** |
| Hub dataset task tags ([worked example](#worked-example-tagging-hub-datasets)), `--base-model fastino/gliner2.5-base-v1 --label-augmentation off`, on `rtx-pro-6000` | single-label choice from a fixed set | 52 | 16,000 × 5 | 17 min | ~$1.50 | top-1 in the owner's tags | 0.320 (always "text-generation") | 0.102 | **0.690** (2 seeds: 0.695 / 0.686) |
Most rows are single, deliberately small runs that test the script, not tuned results. Seed ranges,
out-of-memory history and GPU comparisons are in [GLINER2-NOTES.md](GLINER2-NOTES.md).
### Larger or fixed label sets
For tens of labels scored together (a taxonomy, a fixed tag list), or data in a bucket, see
[Larger or fixed label sets](GLINER2-NOTES.md#larger-or-fixed-label-sets) in the notes:
`--labels-file`, `--label-augmentation off`, local `--train-file`/`--eval-file` and `--export-predictions`.
### Worked example: tagging Hub datasets
A GLiNER2.5-base model fine-tuned with this script on 16,000 Hub datasets suggests task tags for
a dataset from its column names and first row, among the 52 tags the Hub offers. Its first
suggestion matches one of the owner's tags 69% of the time on 3,000 newer datasets from owners it
never saw. Owners' tags are a noisy target: in a hand-checked sample, about 1 in 10 datasets was
missing a tag that fits.
[Try the demo](https://huggingface.co/spaces/davanstrien/hub-task-tagger): paste a dataset id and
see the suggested tags, the owner's tags and the exact text the model read. On a free 2-vCPU Space
one prediction takes about 0.7–1 s.
[Model](https://huggingface.co/davanstrien/hub-task-tagger-gliner2.5-base) · [Notes on how it was trained](GLINER2-NOTES.md#a-larger-label-set-52-hub-task-tags)
### Good to know
- **Single-label and multi-label** are auto-detected from the label column; repeat `--label-column` to train several tasks in one model.
- **Label names are part of the prompt.** Real names (`Fiction`, `Sports`) work; integer codes make zero-shot meaningless.
- **Out of GPU memory, it restarts at a smaller batch size** and stops before pushing if even batch size 1 fails. Many labels or long texts want an `a10g-small`.
- **Always pass `--timeout`.** Older `hf` CLIs ignore the scripts' `[tool.hf-jobs]` header and stop the Job after 30 minutes.
More behaviour details, tested commands, findings and dead ends: [GLINER2-NOTES.md](GLINER2-NOTES.md).
## Few-shot with SetFit (`train-setfit.py`)
An alternative when you have only a handful of labelled examples per class (8-64) and want a sentence-transformer model.
Trains a [SetFit](https://huggingface.co/docs/setfit) classifier from a handful of labelled
examples per class. SetFit finetunes a sentence-transformer body on contrastive pairs, then fits a
logistic regression head on the resulting embeddings.
**Runs on CPU or GPU.** CPU is practical for small few-shot experiments. Use a GPU for faster
training, particularly with larger models, longer texts or more classes. The same model and
training settings work on either; the recipe uses the available accelerator automatically.
```bash
# 8 labels per class, on CPU
hf jobs uv run --flavor cpu-basic --timeout 20m --secrets HF_TOKEN \
https://huggingface.co/datasets/uv-scripts/classification/raw/main/train-setfit.py \
fancyzhx/ag_news username/ag-news-setfit --num-samples 8
# Same model and training settings on a GPU for faster training
hf jobs uv run --flavor t4-small --timeout 20m --secrets HF_TOKEN \
https://huggingface.co/datasets/uv-scripts/classification/raw/main/train-setfit.py \
fancyzhx/ag_news username/ag-news-setfit-gpu --num-samples 8
```
### Measured
8 labels per class, seed 42, evaluated on each dataset's own held-out split (capped at 500
examples, 1000 for banking77):
| Dataset | Classes | Labels used | Body | Flavor | Training | Accuracy | Macro F1 |
|---|---|---|---|---|---|---|---|
| [`SetFit/enron_spam`](https://huggingface.co/datasets/SetFit/enron_spam) | 2 | 16 | MiniLM-L6 | `cpu-basic` | 78s | 0.924 | 0.924 |
| [`fancyzhx/ag_news`](https://huggingface.co/datasets/fancyzhx/ag_news) | 4 | 32 | MiniLM-L6 | `cpu-basic` | 118s | 0.804 | 0.807 |
| [`legacy-datasets/banking77`](https://huggingface.co/datasets/legacy-datasets/banking77) | 77 | 616 | MiniLM-L6 | `t4-small` | 18s | 0.803 | 0.789 |
| [`dair-ai/emotion`](https://huggingface.co/datasets/dair-ai/emotion) | 6 | 48 | MiniLM-L6 | `cpu-basic` | 216s | 0.370 | 0.325 |
**Single seed each — these do not rank models or predict your dataset.** Few-shot results vary
substantially with which examples happen to get sampled; SetFit's own benchmarks report mean and
standard deviation across ten seeds for exactly this reason. Run your own task before trusting
any of these numbers.
### Good to know
- **Default body**: [`all-MiniLM-L6-v2`](https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2) (22M), chosen for CPU speed. Swap it with `--body-model`; set `--max-seq-length` within its context window and add any task prefix it needs yourself. See [Choosing another body or longer context](SETFIT-NOTES.md#choosing-another-body-or-longer-context).
- **Single-label only.** A multi-label column exits with a pointer to `train-classifier.py`.
- **Beat more than the majority baseline.** Every run reports it, but `emotion` beat it by under 2 points; compare with TF-IDF plus logistic regression or zero-shot on the same rows. See [Compare more than the majority baseline](SETFIT-NOTES.md#compare-more-than-the-majority-baseline).
- **Many classes: watch the pair count.** Pairs grow with the square of the training-set size; at 77 classes use `--sampling-strategy undersampling`. See [Many classes](SETFIT-NOTES.md#many-classes-watch-the-pair-count).
- **It refuses runs projected above `--max-minutes`** (default 60). On `biglam/hansard_speech` (2.7M speeches, 28 parties) it refused to train and produced no score, and [the worked failure](SETFIT-NOTES.md#real-world-data-a-worked-failure) shows why.
- **Load it with `SetFitModel.from_pretrained(repo)`**, not `AutoModelForSequenceClassification`: a SetFit model is a sentence-transformer body plus a scikit-learn head.
Evaluation split, metrics, dropped rows and `--private`: [SETFIT-NOTES.md](SETFIT-NOTES.md#behaviour-details).
## Fine-tune a classifier (`train-classifier.py`)
Fine-tunes a text-classification encoder on any Hub dataset and pushes the trained model
back to the Hub — download, train, evaluate, push, and reload-verify in one job.
- **Default model**: [LiquidAI/LFM2.5-Encoder-350M](https://huggingface.co/LiquidAI/LFM2.5-Encoder-350M) — a bidirectional encoder that LiquidAI reports beats ModernBERT-base on GLUE/SuperGLUE, and handles 8,192-token documents. Any Hub encoder works via `--model` (ModernBERT, BERT, DeBERTa, …).
- **Single-label and multi-label**, auto-detected from the label column (`ClassLabel`/string/int → cross-entropy; list of labels → BCE + per-label threshold tuning).
- **Round-trippable artifacts**: standard architectures produce standard models; encoders without a classification head (like LFM2.5) get a generic mean-pooling head pushed as custom code, so `AutoModelForSequenceClassification.from_pretrained(..., trust_remote_code=True)` always works.
```bash
# single-label (ag_news has a ClassLabel column)
hf jobs uv run --flavor a10g-small --secrets HF_TOKEN \
https://huggingface.co/datasets/uv-scripts/classification/raw/main/train-classifier.py \
fancyzhx/ag_news username/news-classifier
# multi-label (go_emotions has a list-of-labels column)
hf jobs uv run --flavor a10g-small --secrets HF_TOKEN \
https://huggingface.co/datasets/uv-scripts/classification/raw/main/train-classifier.py \
google-research-datasets/go_emotions username/emotion-classifier --label-column labels
```
Key options: `--model`, `--max-length` (512 default; up to 8192 with
`--gradient-checkpointing` and a small `--batch-size` on a10g/a100), `--epochs`, `--lr`,
`--batch-size`, `--max-samples` (smoke runs), `--eval-split` (auto-detects
validation/test, or holds out 10% of train). Run `uv run train-classifier.py --help` for all.
### Worked example: classify dataset cards by task
[`davanstrien/dataset-cards-with-task-categories`](https://huggingface.co/datasets/davanstrien/dataset-cards-with-task-categories)
contains 21k Hub dataset cards (frontmatter stripped) labelled with their `task_categories`
metadata — a real multi-label task over long documents:
```bash
hf jobs uv run --flavor a10g-small --secrets HF_TOKEN \
https://huggingface.co/datasets/uv-scripts/classification/raw/main/train-classifier.py \
davanstrien/dataset-cards-with-task-categories username/dataset-card-task-classifier \
--label-column labels --max-length 1024 --batch-size 8 --grad-accum 2
```
The output model predicts likely task categories from a card's prose — e.g. for suggesting
metadata on datasets that lack it.
### Training a standard encoder instead
`--model answerdotai/ModernBERT-base` (or any encoder with a native classification head)
produces a plain, vLLM-servable model — pair it with
[`uv-scripts/vllm`](https://huggingface.co/datasets/uv-scripts/vllm)'s
`classify-dataset.py` for large-scale batch inference with the model you just trained.
## Fine-tune a Kev decision model (`train-kev.py`)
[Kev](https://github.com/jaredpalmer/kev) is an open decision model by Jared Palmer. It answers
yes/no, multiple-choice and rating questions about a text, with a probability for each answer.
[Kev-4B](https://huggingface.co/jaredpalmer/kev-4b) is a rank-16 LoRA adapter and a small pointer
head on Qwen3.5-4B-Base. This script fine-tunes the released Kev-4B on your own labelled questions.
The script is a thin launcher. It pins one Kev commit and the Kev-4B v1.0 weights, installs Kev,
and runs Kev's own `kev.train`, `kev.benchmark` and `kev.calibrate`. All training and evaluation
code is Kev's. Kev's own recipes run on Modal (the `kev-finetune` skill) or on a GPU you have
([Fine-Tune on Your Own Data](https://github.com/jaredpalmer/kev#fine-tune-on-your-own-data));
use this script when your data and credit are already on the Hub.
First put `train.jsonl` and `development.jsonl` in a bucket folder (the snippet at the end of this
section shows one way). Each line is one record in Kev's format; a record can carry more than one
question:
```json
{"state": "<the text>",
"questions": {"newsgroup": {"type": "choice",
"instructions": "Which newsgroup was this post published in?",
"criteria": {"alt.atheism": null, "comp.graphics": null},
"label": "comp.graphics"}}}
```
Then run:
```bash
hf jobs uv run --flavor a100-large --timeout 1h --secrets HF_TOKEN \
-v hf://buckets/<you>/kev-data:/mnt/kev \
https://huggingface.co/datasets/uv-scripts/classification/raw/main/train-kev.py \
/mnt/kev/newsgroups /mnt/kev/runs/newsgroups-v1
```
The script checks every record and the output folder before it installs anything. The output
folder must be new or empty. It gets `checkpoint/`, `finetuned-development/` and
`baseline-development/` (Kev benchmark reports for the fine-tuned and the released model),
`finetuned-development/calibration.json` (when `kev.calibrate` succeeds) and a `run.json` summary.
Each result is copied to the bucket as soon as it is written. On Jobs, the script refuses an output
folder that is not on a mounted bucket.
### Measured
| Task | Train / test records | Flavor | Time | Cost | Released Kev-4B | Fine-tuned |
|---|---|---|---|---|---|---|
| [`SetFit/20_newsgroups`](https://huggingface.co/datasets/SetFit/20_newsgroups) (20 classes) | 846 / 426 (1,000 / 500 prepared) | `a100-large` | 31 min | about $1.29 | 62.4% | 72.3% |
One seed, 2026-10-30. The paired difference is +9.9 points (95% CI +6.1 to +13.6). Training took
about 25 of the 31 minutes (846 records, 2 epochs).
**The fine-tuned model is overconfident**: mean confidence 0.82 against accuracy 0.72 (ECE 0.100).
The fine-tuned checkpoint runs at temperature 1.0. `calibration.json` reports what one temperature
fitted on your development records would do; nothing in the checkpoint changes. On 20 Newsgroups,
the out-of-fold fit (`workload_oof`) brings ECE to 0.028 and mean confidence to 0.73, with accuracy
unchanged (ECE change -0.072, 95% CI -0.092 to -0.023). To use the fitted value
(`workload_temperature`, 1.52 here), set `KEV_TEMPERATURE=1.52` when you serve or benchmark the
checkpoint with Kev's commands, or pass `LoadOptions(temperature=1.52)` when you load it in Python.
### Good to know
- **Long records are dropped.** `kev.train`'s default context is 384 state tokens (Kev's `--max_state`
raises it; this script does not expose it). `kev.train` drops longer records and prints how many; `kev.benchmark` drops them too and lists them in `rejected.json`. On
20 Newsgroups, after cutting posts at 2,000 characters, this dropped 154 of 1,000 training and 74 of
500 test posts.
- **Raise `--timeout` before you add data or epochs.** Training time grows with records × epochs.
If the Job hits the timeout, the results already copied stay in the bucket.
- **Everything is pinned:** the Kev commit, the Kev-4B weights (by commit, because a tag can move)
and the Qwen3.5 base revision. If the `kev-1.0` tag moves, the script stops with a message.
- **No extra kernels.** Kev's Modal images add `flash-linear-attention` (and, in `modal_app.py`,
`causal-conv1d`).
In our `a100-large` runs they gave no speed-up, so this script leaves them out.
- **Smoke test first:** `--max-steps 5` runs the whole path in about 6 minutes.
- **No 20 Newsgroups checkpoint is published.** `SetFit/20_newsgroups` declares no licence.
<details><summary>Turn a labelled dataset into Kev records (the 20 Newsgroups example)</summary>
```python
import json
from pathlib import Path
from datasets import load_dataset
data = load_dataset("SetFit/20_newsgroups")
labels = sorted(set(data["train"]["label_text"]))
out = Path("newsgroups")
out.mkdir(exist_ok=True)
for split, count, name in [("train", 1000, "train.jsonl"), ("test", 500, "development.jsonl")]:
rows = data[split].filter(lambda row: len(row["text"].strip()) >= 50).shuffle(seed=42).select(range(count))
with (out / name).open("w") as f:
for row in rows:
question = {"type": "choice", "instructions": "Which newsgroup was this post published in?",
"criteria": {label: None for label in labels}, "label": row["label_text"]}
f.write(json.dumps({"state": row["text"].strip()[:2000], "questions": {"newsgroup": question}}) + "\n")
```
Then copy the folder to your bucket: `hf buckets sync ./newsgroups hf://buckets/<you>/kev-data/newsgroups`.
</details>
## Zero-shot classification (`classify-dataset.py`)
Label a dataset with an instruction LLM and no training data. You give a list of labels; the
model picks one per row, and guided decoding (structured outputs) makes sure every answer is one
of your labels. The default model is
[HuggingFaceTB/SmolLM3-3B](https://huggingface.co/HuggingFaceTB/SmolLM3-3B); any instruction
model works via `--model`. The result is your dataset with a new `classification` column.
Two scripts do this. `classify-dataset.py` runs on vLLM and is the one to start with.
`classify-dataset-sglang.py` runs the same task on SGLang, for reasoning models that write
`<think>` traces; its options differ (`--reasoning`, `--save-reasoning`, `--batch-size`,
`--grammar-backend`), so check its `--help`.
### Quick start
A GPU is required. Run on Jobs with the vLLM image:
```bash
hf jobs uv run --flavor l4x1 --image vllm/vllm-openai:v0.32.0 --secrets HF_TOKEN \
https://huggingface.co/datasets/uv-scripts/classification/raw/main/classify-dataset.py \
--input-dataset stanfordnlp/imdb \
--column text \
--labels "positive,negative" \
--output-dataset username/imdb-classified \
--max-samples 100 --shuffle
```
`--max-samples` with `--shuffle` takes a random sample, which matters for datasets sorted by
date or label. Drop both to label the whole split.
### With reasoning and label descriptions
```bash
hf jobs uv run --flavor l4x1 --image vllm/vllm-openai:v0.32.0 --secrets HF_TOKEN \
https://huggingface.co/datasets/uv-scripts/classification/raw/main/classify-dataset.py \
--input-dataset user/support-tickets \
--column content \
--labels "bug,feature_request,question,other" \
--label-descriptions "bug:code or product not working as expected,feature_request:asking for new functionality,question:seeking help or clarification,other:general comments or feedback" \
--enable-reasoning \
--output-dataset username/tickets-classified
```
With `--enable-reasoning` the model thinks step by step before it answers, and the output also
has `reasoning` and `parsing_success` columns. Reasoning mode turns off structured outputs: the
model must end with `{"label": "..."}`, and rows where that cannot be parsed are marked in
`parsing_success`. It is slower, but you can read why each label was chosen.
### Options
| Option | What it does |
|---|---|
| `--model` | Model to use (default `HuggingFaceTB/SmolLM3-3B`) |
| `--label-descriptions` | `label:description,...` pairs that tell the model what each label means |
| `--enable-reasoning` | Think before answering; adds `reasoning` and `parsing_success` columns |
| `--split` | Split to process (default `train`) |
| `--max-samples` | Label only the first N rows (or N random rows with `--shuffle`) |
| `--shuffle`, `--shuffle-seed` | Shuffle before `--max-samples` (seed default 42) |
Run `uv run classify-dataset.py --help` for all options (no GPU or vLLM needed).
### Good to know
- **Speed**: about 50-100 texts/second for SmolLM3-3B on an A10, and 20-50 for 7B models.
`l4x1` is a good start; use `a10g-large` or larger for 7B+ models or out-of-memory errors.
- **Text handling**: texts shorter than 3 characters and empty values are skipped; texts are
truncated to 4,000 characters.
- **Label names matter.** Use clear, distinct names, add `--label-descriptions` when names are
ambiguous, and try a larger model for nuanced tasks.
- **vLLM version**: on Jobs, vLLM, PyTorch and Transformers come from the `vllm/vllm-openai`
image, so the script header does not list them. For a local GPU run (Linux + CUDA), add vLLM
yourself: `uv run --with vllm classify-dataset.py ...`. The script needs vLLM 0.12 or newer
(`StructuredOutputsParams`).
|