# arcee-ai/DistillKit

An Open Source Toolkit For LLM Distillation

Repository: https://github.com/arcee-ai/DistillKit
Canonical: https://ross.abutalabs.com/products/distillkit
Language: Python
License: Apache-2.0
License Family: permissive
Last push: 2026-05-12T20:10:14+00:00

## Health v2 (maintenance only)
Score: 60/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 82, release rhythm 35, longevity 54
- inputs: {"age_days": 763, "days_push": 113, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: no_releases
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 1047, forks 139 (observed 2026-08-28T04:03:21.940657+00:00)

## What it is
DistillKit is an open-source Python toolkit for knowledge distillation of large language models, supporting both online and offline distillation workflows. It includes an advanced logit compression system (polynomial approximation, quantization, and bit-packing) that makes large-scale offline distillation practical.

## Use cases
- distill a large teacher LLM into a smaller student model
- run offline distillation from pre-captured teacher logits
- compress teacher logit datasets for large-scale distillation
- train a student model with KL divergence and other distillation losses
- capture teacher model outputs for later distillation training

## When to choose
- you need to distill an LLM at scale using pre-captured teacher outputs
- you want battle-tested tooling that has powered production model releases
- you want flexible, composable distillation loss functions
- you already use the HuggingFace Transformers/TRL/Accelerate ecosystem

## When to avoid
- you need general-purpose LLM fine-tuning rather than distillation
- you want a no-code or fully managed training service
- you lack GPU resources for model training
- you need distillation for non-language modalities

## Facets
- artifact type: library
- maturity: active
- function: llm-training, machine-learning, deep-learning, gpu-computing
- domain: large-language-models, machine-learning, deep-learning
- platform: python
- tags: knowledge-distillation, logit-compression, huggingface, teacher-student-training, offline-distillation, gpu, linux

## Member repositories
- arcee-ai/DistillKit (main) score 60

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:03:21.940657+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T07:01:05.060650+00:00, confidence not recorded.
  - readme: https://github.com/arcee-ai/DistillKit (fetched 2026-08-28T04:03:21.940657+00:00, sha 93c6e0772ce7)
- Data as of 2026-08-30T08:39:29.467469+00:00.