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.

Results

Two tasks with Qwen/Qwen2.5-1.5B-Instruct as the zero-shot teacher, and Banking77 also with Qwen/Qwen2.5-7B-Instruct; the student is always sentence-transformers/all-MiniLM-L6-v2. Quality is measured on gold-labelled test texts that neither model trains on. Every table is written by the task's run.sh from its saved results*.json file in examples/; no number is typed by hand. The teacher's latency includes scoring every surface form of every label, which costs most on Banking77 with its many labels, so there the GPU teacher row is the fairer reference for speed.

Banking77: the hard case

Intent detection with 77 fine-grained classes (Banking77, CC BY 4.0), reproduced by examples/banking77/run.sh.

Teacher only, no gold labels:

Task banking77, 3080 gold-labelled test texts. Teacher: Qwen/Qwen2.5-1.5B-Instruct. Student: sentence-transformers/all-MiniLM-L6-v2. The student learns only from the teacher's soft labels on 10003 unlabelled training texts. 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 scores one text at a time, so it is timed at batch 1 and its throughput is 1 / latency. Speed-up is relative to the teacher, zero-shot on cpu.

With 8 gold labels per class, and the student trained on all gold labels as an upper bound:

Task banking77, 3080 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 10003 training texts, plus cross-entropy on 8 gold labels per class (616 texts, weight 1.0). 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. 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 scores one text at a time, so it is timed at batch 1 and its throughput is 1 / latency. Speed-up is relative to the teacher, zero-shot on cpu.

With a stronger teacher, Qwen/Qwen2.5-7B-Instruct loaded in 4 bit (timed on the GPU only: 4-bit loading needs CUDA), teacher only:

Task banking77, 3080 gold-labelled test texts. Teacher: Qwen/Qwen2.5-7B-Instruct. Student: sentence-transformers/all-MiniLM-L6-v2. The student learns only from the teacher's soft labels on 10003 unlabelled training texts. 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 scores one text at a time, so it is timed at batch 1 and its throughput is 1 / latency. Speed-up is relative to the teacher, zero-shot, 4-bit on cuda.

The same teacher with 8 gold labels per class and the upper bound:

Task banking77, 3080 gold-labelled test texts. Teacher: Qwen/Qwen2.5-7B-Instruct. Student: sentence-transformers/all-MiniLM-L6-v2. 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 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. 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 scores one text at a time, so it is timed at batch 1 and its throughput is 1 / latency. Speed-up is relative to the teacher, zero-shot, 4-bit on cuda.

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.

Task financial-tweets, 2388 gold-labelled test texts. Teacher: Qwen/Qwen2.5-1.5B-Instruct. Student: sentence-transformers/all-MiniLM-L6-v2. The student learns only from the teacher's soft labels on 9543 unlabelled training texts. The upper bound is the same student trained with cross-entropy on all 9543 gold training labels and no teacher; its speed equals the student's. 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 scores one text at a time, so it is timed at batch 1 and its throughput is 1 / latency. Speed-up is relative to the teacher, zero-shot on cpu.

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, and --cache-dir (runs that share it label each text only once). See tinydistill <command> --help.

Few labels. With --gold-per-class K, the student also sees the gold labels of K randomly chosen training examples per class, through a cross-entropy term (weight --ce-weight, default 1) added to the distillation loss. --upper-bound additionally trains the same student on all gold training labels without the teacher, and the results table gets an upper-bound row: how far the distilled student is from fully supervised training.

From Python:

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 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}

  Message: {text}

  Reply with the intent exactly as written in the list.
labels:
  - activate_my_card    # a bare id: its name is derived, "activate my card"
  - card_arrival
  # ...

{labels} (required) becomes a bulleted list of the label names, and {text} is the input. 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. Tests check the scores against a plain forward pass per continuation.

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.
• 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), one text at a time.
• Single-label classification only. 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.

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.