tinydistill

Most of a large model's quality on one task, many times faster, on a CPU.

tinydistill turns an LLM classification prompt into a small encoder classifier by knowledge distillation. You describe the task in a short YAML file (labels, how to say each label, an instruction prompt) and give it unlabelled texts. An instruction-tuned LLM labels them with soft targets, a small encoder learns to imitate those targets, and the result is exported to ONNX (optionally int8) for fast CPU inference.

It continues my bachelor's project on inference optimisation by distilling LLMs.

flowchart LR
    A["Task prompt +<br/>unlabelled texts"] --> B["LLM teacher(s)<br/>Qwen · Mistral · Phi · ..."]
    S["Synthetic texts<br/>(synth)"] -.-> B
    V["vLLM server"] -.-> B
    K["A few gold labels<br/>per class"] -.-> B
    K -.-> C
    B -- "soft labels<br/>(ensembled)" --> C["Small encoder<br/>student"]
    C --> D["ONNX / int8<br/>export"]
    D --> E["Fast CPU<br/>inference"]
    D --> W["Browser · server ·<br/>Hugging Face Hub"]
    E -.-> F["Cascade: unsure<br/>texts to the LLM"]

Each step is one command: tinydistill label (teacher), train (student), export (ONNX / int8) and predict (inference), or tinydistill run for all of them. Dashed parts are optional: synthetic texts when there is no unlabelled data, a vLLM server as the teacher, a few gold labels for the teacher's prompt and the student's loss, and a cascade that hands the student's least confident texts to the LLM. The exported model runs on a CPU, in a browser, behind tinydistill serve, or from the Hugging Face Hub.

Results

Teachers are Qwen/Qwen2.5-1.5B-Instruct and Qwen/Qwen2.5-7B-Instruct (4 bit), zero-shot or with 8 gold examples per class in the prompt; the student is sentence-transformers/all-MiniLM-L6-v2 unless a table says otherwise. Quality is measured on gold-labelled test texts that no model trains on, and students are trained with 3 seeds (mean ± standard deviation). Every table below is filled by the task's run.sh from its saved results files in examples/; the text only quotes them.

Try it: the live demo runs the Banking77 student (7B 8-shot teacher, int8, 23 MB) in your browser and shows what happens inside it: the word pieces, what each layer would answer, where the classifier looks, which words matter, and the teacher's soft labels next to the student's.

Banking77: the hard case

Intent detection with 77 fine-grained classes (Banking77, CC BY 4.0), reproduced by examples/banking77/run.sh. One row per teacher; the student columns are the deployed int8 model:

All latencies on the CPU with 2 threads.

The student follows its teacher, so the teacher sets the ceiling: a bigger model and 8 examples per class each help a lot. The full table for the best teacher, with the student trained on the same 616 gold labels alone (baseline) and on all 10,003 gold labels (upper bound), and the cascade:

Task banking77, 3080 gold-labelled test texts. Teacher: Qwen/Qwen2.5-7B-Instruct. Student: sentence-transformers/all-MiniLM-L6-v2. The teacher's prompt shows 8 gold training examples per class (616 texts). The student learns from the teacher's soft labels on all 10003 training texts, plus cross-entropy on 8 gold labels per class (616 texts, weight 1.0). The baseline is the same student trained with cross-entropy on those 616 gold labels only, without the teacher. The upper bound is the same student trained with cross-entropy on all 10003 gold training labels and no teacher; its speed equals the student's. Values with ± are the mean and standard deviation over 3 training seeds; the ONNX rows are the exported first-seed model. CPU rows use 2 threads. Latency is the median over repeated runs. Size is the teacher's weights as loaded and the students' model files. The teacher packs several texts into one forward pass; it is timed at batch 32 only when that takes under three minutes, otherwise its throughput is 1 / batch-1 latency. Speed-up is relative to the teacher, 8-shot, 4-bit on cuda.

Cascade: student first, unsure texts to the teacher

The student answers every text and passes its least confident ones to the teacher. Thresholds are set on the test texts to hit each share; in deployment, set them on a validation split. Latency is the student's batch-1 latency plus the share times the batch-1 latency of the teacher, 8-shot, 4-bit on cuda.

The distilled student is far better than the same student trained on the few gold labels alone, and runs on a CPU in about a millisecond. Sending its least confident texts to the teacher closes most of the remaining gap to the teacher.

Other students, the same teacher and soft labels. Students train for 5 epochs unless the name says otherwise. static is the static-embedding student (--student static: token vectors from MiniLM-L6, averaged, no attention), trained for 20 epochs at learning rate 1e-3:

All latencies on the CPU with 2 threads.

Training longer matters more than a bigger student: with 20 epochs instead of 5, the 23M-parameter MiniLM-L6 student is as accurate as the 110M-parameter mpnet, at a sixth of its latency, and more accurate than its own 7B teacher. Learning from the teacher's soft labels on 10,000 texts, plus the 616 gold labels, smooths out some of the teacher's individual mistakes. The static student is about a point behind the MiniLM students trained as long, at a tenth of a millisecond per text and 8 MB. Ten epochs get most of the way (MiniLM-L6: 0.808), so 10 is the default from 0.3.1. On the easier financial-tweets task, 20 epochs give the same accuracy as 5 (financial-tweets-7b-20-epochs.md). The other tables on this page were trained for 5 epochs, the default until 0.3.1.

Financial tweets: an easier task

Sentiment of financial news tweets, 3 classes (Twitter Financial News Sentiment, MIT licence per the dataset card), reproduced by examples/financial-tweets/run.sh. The validation split is the test set. The rows include an ensemble of three model families (Qwen2.5-7B, Mistral-7B and Phi-3.5-mini, all 4 bit) and a student trained only on texts the teacher wrote itself (tinydistill synth, no real training texts):

All latencies on the CPU with 2 threads.

Here the 7B teacher is close to the student trained on all gold labels, and the student is close to the teacher. The ensemble adds a little. Few-shot examples hurt on this skewed task (see Limitations). The student that saw only about 1,500 texts the teacher wrote itself is far behind the one trained on real tweets: generated texts are much less varied than real ones, so synth is a way to start without data, not a replacement for it. Full tables per run are in examples/financial-tweets/.

GoEmotions: multi-label

Reddit comments with any of 27 emotions or neutral (GoEmotions, Apache 2.0), reproduced by examples/go-emotions/run.sh. The teacher answers one yes/no question per label; accuracy is the share of texts with exactly the right set of labels.

Task go-emotions, 5427 gold-labelled test texts. Teacher: Qwen/Qwen2.5-1.5B-Instruct. Student: sentence-transformers/all-MiniLM-L6-v2. The student learns from the teacher's soft labels on all 10000 training texts, plus cross-entropy on 8 gold labels per class (223 texts, weight 1.0). The baseline is the same student trained with cross-entropy on those 223 gold labels only, without the teacher. The upper bound is the same student trained with cross-entropy on all 10000 gold training labels and no teacher; its speed equals the student's. Values with ± are the mean and standard deviation over 3 training seeds; the ONNX rows are the exported first-seed model. CPU rows use 2 threads. Latency is the median over repeated runs. Size is the teacher's weights as loaded and the students' model files. The teacher packs several texts into one forward pass; it is timed at batch 32 only when that takes under three minutes, otherwise its throughput is 1 / batch-1 latency. Speed-up is relative to the teacher, zero-shot on cpu.

This is a hard setting for a 1.5B teacher: it answers yes far too often, so exact-set accuracy is near zero, and the student inherits that. The run shows that multi-label tasks work end to end, not that a small teacher is good at them; use a larger teacher.

A vLLM teacher agrees with the local one

examples/vllm-smoke.sh labels 300 financial tweets with Qwen/Qwen2.5-1.5B-Instruct locally and through a vLLM server:

Qwen2.5-1.5B-Instruct scoring 300 financial tweets locally (bf16, tinydistill) and through vLLM (prompt log-probabilities); label scores are log-probabilities, so small differences are numerical.

Quickstart

pip install "tinydistill[hf]"
curl -LO https://raw.githubusercontent.com/Andreas-Hum/tinydistill/main/examples/banking77/task.yaml
tinydistill run task.yaml --data hf:PolyAI/banking77:train --test hf:PolyAI/banking77:test \
  --out runs/banking77 --int8
tinydistill predict runs/banking77/onnx "I still haven't received my new card"

Labelling the full dataset with the default teacher wants a GPU. For a quick look on a CPU, add --teacher Qwen/Qwen2.5-0.5B-Instruct --limit 200 --test-limit 200. --limit takes the first rows, and the Banking77 splits are sorted by intent, so a limited run only sees a few intents.

run does every step. Each step also has its own command, and they all read and write the run directory:

Data can be .csv, .jsonl or hf:<dataset>[:<config>]:<split> (needs the hf extra). Useful options: --teacher (any Hugging Face causal LM), --load-4bit (bitsandbytes, CUDA only: pip install "tinydistill[4bit]"), --student (any Hugging Face encoder), --epochs, --lr, --device, --threads, --seed, and --cache-dir (runs that share it label each text only once, and a run whose texts are all cached does not load the teacher at all). See tinydistill <command> --help.

Long steps print their progress, speed and time left every few seconds and write it to RUN/status.json; tinydistill status RUN reads it from another terminal and tells you whether the process is still alive.

Static student. --student static (or static:<encoder>) trains a student with no attention at all: one vector per token, averaged, then a linear layer, started from the encoder's own token representations (as in Model2Vec). It is much smaller and faster than a transformer student. It ignores word order, so it can miss meaning that depends on it (negation, "A rather than B"); on Banking77 it came within about a point of MiniLM students trained as long (see the results). It exports to ONNX and runs with predict and serve like any other student, but not in the browser page. Train it longer and at a higher learning rate, such as --epochs 20 --lr 1e-3.

Choosing the teacher

--teacher takes any Hugging Face causal LM with a chat template; the tests check the scoring against a plain forward pass for Llama, Mistral, Phi, Gemma 2 and GPT-2 style models as well as Qwen. Several teachers, comma-separated, are an ensemble: each scores every text (one model in memory at a time) and the soft labels are the mean of their label probabilities. tinydistill compare then puts the runs side by side.

A vLLM server can be the teacher instead of a local model, so a large model on a GPU cluster does the labelling: start it with vllm serve MODEL --enable-prefix-caching and pass --vllm-url http://host:8000/v1 with --teacher MODEL. Only the tokenizer is loaded locally; every continuation is scored exactly from the server's prompt log-probabilities, so the soft labels are the same as with the local model.

Many labels. With hundreds or thousands of labels, scoring every one for every text gets slow. label --shortlist K first compares each text with the label names (and descriptions) using a small sentence-embedding model (--shortlist-model, MiniLM by default) and has the teacher score only the K most similar labels; the others get probability zero. eval reports the shortlist recall: how often the gold label was among the K. If it is low, raise K.

Tuning the setup. tinydistill tune TASK.yaml --data D --teacher a,b --few-shot 0,8 --prompt other.yaml scores every combination of teacher, number of few-shot examples and prompt on a dev split of gold-labelled training texts (--dev-size, default 200; never the test set) and writes tune.md and tune.json (--out), best first, with the label command for the winner.

No unlabelled data?

tinydistill synth TASK.yaml --per-label 200 --out texts.jsonl asks the teacher to write example inputs for each label, from the task's own instruction. This is the only place where the teacher generates text, and only inputs: the file has no label field, and the texts are labelled by the usual scoring when you pass them as --data. Generated texts are less varied than real ones; use real texts when you have them.

Deploying: cascade, server, browser

Cascade. eval adds a table of what happens when the student answers and passes the texts it is least sure about to the teacher: accuracy and mean latency for 0 % to 100 % of the texts sent on. At inference, tinydistill predict RUN/onnx --cascade RUN --threshold T (and serve --cascade) does this for real: texts whose top student probability is below T go to the run's teacher. Choose T on a validation split.

Server. tinydistill serve RUN/onnx answers POST /predict with {"texts": [...]} and GET /health, using only the standard library. It listens on 127.0.0.1 unless you pass --host.

Browser. tinydistill web RUN/onnx --out site writes a static page plus the model in the layout transformers.js expects; the int8 model then runs in the visitor's browser with no server at all. Below the answer, an "Inside the model" view shows each step of the inference (tokens, the answer after each layer, attention, probabilities, which words matter, the teacher's soft labels for the examples and the cascade), so the page also explains distillation. Serve the folder with any static host (GitHub Pages works).

A few gold labels

If some training texts have gold labels, they can help in three places. All of them use the same K randomly chosen examples per class (fixed by --seed):

• label --few-shot K shows the examples in the teacher's prompt, in place of an {examples} line in the task's prompt. The prompt prefix is computed once, so even hundreds of examples cost little per text.
• train --gold-per-class K adds a cross-entropy term on their gold labels (weight --ce-weight, default 1) to the distillation loss.
• train --gold-only also trains the student on those labels alone, without the teacher, as a baseline row: what the teacher adds over the few labels.

train --upper-bound trains the same student on all gold training labels, again without the teacher: how far distillation is from full supervision. train --seeds N repeats every student with N training seeds, and the results show the mean and standard deviation.

Improving a run

• tinydistill analyse RUN writes RUN/analysis.html: precision, recall and F1 of every label for the teacher and the student, the most common confusions with example texts, and texts where the two disagree. Labels the teacher gets wrong usually need a clearer name, a description, or merging.
• tinydistill suggest RUN --n 200 --out to_label.jsonl picks the training texts most worth labelling by hand: where the student and the teacher disagree, then where the teacher is least sure, spread over labels. Fill in the empty label field and pass the file back with train RUN --gold-file to_label.jsonl; those labels join the cross-entropy term.
• predict --cascade RUN --log cascade.jsonl (and serve --cascade ... --log) records every text the cascade passed to the teacher, with the teacher's scores. tinydistill refresh RUN --log cascade.jsonl --out RUN2 adds them to the training texts and trains, exports and evaluates a new student on the same test set, so the student learns the inputs it found hard in production.

From Python

The exported model:

from tinydistill import Predictor

clf = Predictor("runs/banking77/onnx")  # uses model.int8.onnx when present
clf.predict(["How do I top up my card?"])  # list of label names
clf.predict_proba(["How do I top up my card?"])  # numpy array, one row per text

The whole pipeline, without the command line:

from pathlib import Path

import torch
from tinydistill import Predictor, export_onnx, load_task, load_teacher, read_examples, train_student

task = load_task("task.yaml")
texts = [e.text for e in read_examples("hf:PolyAI/banking77:train", task, limit=1000)]
teacher = load_teacher("Qwen/Qwen2.5-1.5B-Instruct", task, device="cuda")
soft_targets = torch.tensor(teacher.score_many(texts)).softmax(-1)  # one row of label scores per text
train_student(texts, soft_targets, task.labels, Path("student"))
export_onnx(Path("student"), Path("student-onnx"), int8=True)
Predictor("student-onnx").predict(["Where is my card?"])

The task file

name: banking77
text_field: text        # column holding the input text
label_field: label      # column holding gold labels (ids, or integers indexing `labels`)
prompt: |-
  You are classifying messages that customers sent to their bank's support chat.
  Choose the single intent from this list that best matches the message:
  {labels}
  {examples}

  Message: {text}

  Reply with the intent exactly as written in the list.
surface_forms: [lowercase, capitalised]   # optional, see below
labels:
  - activate_my_card    # a bare id: its name is derived, "activate my card"
  - card_arrival
  # ...

For multi-label tasks, where a text can have any number of labels, add multi_label: true and ask about one label at a time, with {label} after {text}:

multi_label: true
prompt: |
  Review: {text}
  Does the review talk about {label}? Answer yes or no.
labels: [price, delivery, product_quality]

The teacher answers every label's question in the same forward pass, and a label's score is log P(yes) − log P(no) over the surface forms of both answers (answers: [yes, no] changes them). The student gets a multi-label head trained against the teacher's per-label probabilities, and the results show exact-match accuracy, macro-F1 and micro-F1, with 0.5 as the cut-off. Gold labels can be lists or strings like a|b. Few-shot examples and the cascade are for single-label tasks.

For labels with an order, such as 1 to 5 stars, add ordinal: true: the results then also show the mean absolute error in label steps and the share of predictions at most one step off.

{labels} (required) becomes a bulleted list of the label names, {text} is the input, and an optional {examples} line holds few-shot examples (it disappears when there are none). The name is what the teacher scores. Each label can be a bare id as above, id: name, or carry a description that is shown in the list (- bullish: positive for the stock) but not scored:

labels:
  bearish:
    description: negative for the stock, company or market
  lost_card: lost or stolen card

How the label scoring works

The teacher never generates free text. For an input x, the prompt is rendered with the model's chat template (user message = the instruction with the label list, generation prompt for the assistant). A continuation is one way of writing a label followed by the end-of-turn token, c_1 ... c_m, with log-probability

log p(c | prompt(x)) = sum_j log p(c_j | prompt(x), c_1 ... c_(j-1))

Multi-token labels are handled exactly, and the end-of-turn token keeps a label that is a prefix of another (card versus card arrival) from winning by stopping early. Models do not always write a label the way it is listed: Qwen2.5 tends to answer a sentiment question with "Bullish" even when the list says "bullish". So each label is scored over its common surface forms, the name as written, lowercase, Capitalised and UPPERCASE, each with and without a leading space (forms that tokenise identically count once):

score(label) = log sum over forms c of p(c | prompt(x))

This is verbaliser coverage: it only counts more ways of writing the same answer, and there is no calibration or prior correction. All forms share the prefix tree below, but with many labels they still add up; a task file can list fewer (surface_forms: [lowercase, capitalised] drops UPPERCASE, which tokenises into many pieces, and surface_forms: [] scores the name as written only). A softmax over the label scores gives the soft targets. The prompt lists each label in its primary form, the name as written. The scores are cached as JSON lines keyed by a hash of the text, in a file named after a fingerprint of the teacher, prompt and continuations, so an interrupted run resumes and a changed prompt never reuses old scores.

To keep this fast, the part of the prompt before the text is run once and its key/value cache is kept. Each text then needs a single forward pass that covers the text and all labels together: the tokens of all continuations (every surface form of every label) form a prefix tree, so continuations that start alike, such as card arrival and card linking, share nodes, and a 4D attention mask lets each tree node attend only to the prompt, the text and its own ancestors. There is no padding and no per-label copy of the cache. When a task has few labels and short texts, several texts are packed into one forward pass the same way, each attending only to the prompt and itself (--teacher-tokens, default 2048 tokens per pass). Tests check the scores against a plain forward pass per continuation, and packed against unpacked scores.

The student minimises KL(teacher || student) over the training texts, plus --ce-weight times the cross-entropy on the K gold labels per class chosen by --gold-per-class, if any (standard knowledge distillation, Hinton et al., 2015).

Limitations

• The teacher caps the student. The student learns to imitate the teacher, mistakes included. A small LLM choosing zero-shot among many fine-grained labels can be weak: Qwen2.5-0.5B-Instruct, for example, often confuses related Banking77 intents. A bigger teacher (try --load-4bit with a 7B model) or a better prompt usually helps more than anything on the student side.
• Scores are sums, not averages. Longer verbalisations get lower scores, so very uneven label lengths bias the teacher. Keep verbalisations short and of similar length.
• Only case and spacing variants are counted. Plurals, synonyms or rephrasings of a label are not; keep label names short and close to how the teacher would answer.
• A too-small teacher can collapse. Qwen2.5-0.5B-Instruct, for example, puts nearly every text into one or two labels on the financial-tweets task, whatever the scoring; use 1.5B or larger.
• Few-shot examples can hurt. K examples per class tell the teacher that the classes are equally common, and the last examples in the prompt pull it towards their labels. On a skewed task the few-shot teacher can be worse than the zero-shot one; compare both (tinydistill compare) before relying on it.
• Prompt and label order matter. Zero-shot LLMs are sensitive to wording and to the order of the options.
• Teacher cost grows with labels. Every text is scored against every surface form of every label (shared prefixes are computed once). With very many labels use --shortlist K; a gold label outside the shortlist cannot be learned from the teacher for that text.
• Multi-label tasks ask one question per label, so they cost one forward pass per label and text. Inputs are truncated to 128 tokens by default (--max-length).
• Timings depend on the machine. The results table states the threads and devices used; rerun eval on your own hardware.
• ONNX export uses PyTorch's TorchScript-based exporter.

Development

python -m venv .venv && . .venv/bin/activate
pip install torch --index-url https://download.pytorch.org/whl/cpu
pip install -e ".[dev]"
ruff check . && ruff format --check . && pytest

The tests run on a CPU in a few seconds with tiny random models from hf-internal-testing. They check the label scores against a direct computation, the distillation loss and metrics against hand-computed values, cache resumption, that the ONNX export reproduces the PyTorch logits, and the whole CLI end to end.

See CONTRIBUTING.md for how to propose changes, SECURITY.md for reporting vulnerabilities and CODE_OF_CONDUCT.md for the code of conduct.

References

• G. Hinton, O. Vinyals, J. Dean. Distilling the Knowledge in a Neural Network. arXiv:1503.02531, 2015.
• T. Schick, H. Schütze. Exploiting Cloze Questions for Few Shot Text Classification and Natural Language Inference. EACL, 2021. (Verbalisers.)
• I. Casanueva, T. Temčinas, D. Gerz, M. Henderson, I. Vulić. Efficient Intent Detection with Dual Sentence Encoders. NLP4ConvAI Workshop, 2020. (Banking77.)
• W. Wang, F. Wei, L. Dong, H. Bao, N. Yang, M. Zhou. MiniLM: Deep Self-Attention Distillation for Task-Agnostic Compression of Pre-Trained Transformers. NeurIPS, 2020.
• N. Reimers, I. Gurevych. Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks. EMNLP, 2019.
• Qwen Team. Qwen2.5 Technical Report. arXiv:2412.15115, 2024.

Licence

MIT. The Banking77 dataset is CC BY 4.0 (PolyAI); models are downloaded from the Hugging Face Hub under their own licences.