distill-anything 0.1.0__tar.gz

distill_anything-0.1.0/.gitignore

@@ -0,0 +1,13 @@
+__pycache__/
+*.py[cod]
+*.egg-info/
+.venv/
+dist/
+build/
+runs/
+.pytest_cache/
+.ruff_cache/
+.DS_Store
+*.log
+headroom_readme.md
+.env

distill_anything-0.1.0/CONTRIBUTING.md

@@ -0,0 +1,45 @@
+# Contributing to Distill Anything
+
+Thanks for helping build the open-source distillation lifecycle. PRs of all sizes welcome.
+
+## Dev setup
+
+```bash
+git clone https://github.com/AIAnytime/distillanything && cd distillanything
+uv venv && uv pip install -e ".[dev]"
+```
+
+## Before you open a PR
+
+```bash
+ruff check distillanything tests   # lint
+pytest                             # full suite — offline, no GPU, no API keys, no downloads
+distill smoke                      # end-to-end pipeline check on your hardware
+```
+
+The test suite is deliberately hermetic: tiny random models from `distillanything/testing.py`
+and fake judges stand in for real models and APIs. **Keep it that way** — a test that needs
+a network connection, an API key, or a model download will not be merged. If your feature
+touches an API teacher, test the logic around the call with a `Teacher` fake (see
+`tests/test_judge.py` for examples).
+
+## What makes a good contribution
+
+- **Roadmap items** (see README) are pre-approved directions — open an issue to claim one.
+- **New teachers**: subclass `Teacher`, register a spec prefix in `teachers/registry.py`,
+  keep SDK imports lazy so the core install stays light.
+- **New losses**: pure functions in `losses/`, operating on `[B, S, V]` logits with the
+  `-100` label-mask convention, plus property tests (non-negativity, zero-at-identity,
+  masking, backprop) like `tests/test_losses.py`.
+- **Recipes**: sized so they run on a 16GB laptop, or clearly labeled otherwise.
+
+## Style
+
+- `ruff` (line length 110) — CI enforces it.
+- Comments explain constraints the code can't show, not what the next line does.
+- User-facing strings and errors should say what to do next, not just what went wrong.
+
+## Releases
+
+Maintainers cut releases from `main`. Version lives in `pyproject.toml` and
+`distillanything/__init__.py`.

distill_anything-0.1.0/LICENSE

@@ -0,0 +1,201 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright 2026 Distill Anything contributors
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

distill_anything-0.1.0/PKG-INFO

@@ -0,0 +1,329 @@
+Metadata-Version: 2.4
+Name: distill-anything
+Version: 0.1.0
+Summary: The open-source lifecycle framework for distilling any model: generate data from teachers, distill students, evaluate, and benchmark.
+Project-URL: Homepage, https://github.com/AIAnytime/distillanything
+Project-URL: Repository, https://github.com/AIAnytime/distillanything
+Project-URL: Issues, https://github.com/AIAnytime/distillanything/issues
+Author: Distill Anything contributors
+License-Expression: Apache-2.0
+License-File: LICENSE
+Keywords: distillation,knowledge-distillation,llm,synthetic-data,training
+Classifier: Development Status :: 3 - Alpha
+Classifier: Intended Audience :: Developers
+Classifier: Intended Audience :: Science/Research
+Classifier: Programming Language :: Python :: 3
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Requires-Python: >=3.10
+Requires-Dist: pydantic>=2.5
+Requires-Dist: pyyaml>=6.0
+Requires-Dist: rich>=13.0
+Requires-Dist: torch>=2.2
+Requires-Dist: transformers>=4.44
+Requires-Dist: typer>=0.12
+Provides-Extra: anthropic
+Requires-Dist: anthropic>=0.40; extra == 'anthropic'
+Provides-Extra: api
+Requires-Dist: anthropic>=0.40; extra == 'api'
+Requires-Dist: openai>=1.40; extra == 'api'
+Provides-Extra: dev
+Requires-Dist: pytest>=8.0; extra == 'dev'
+Requires-Dist: ruff>=0.6; extra == 'dev'
+Provides-Extra: openai
+Requires-Dist: openai>=1.40; extra == 'openai'
+Description-Content-Type: text/markdown
+
+<div align="center"><pre>
+  ██████╗ ██╗███████╗████████╗██╗██╗     ██╗
+  ██╔══██╗██║██╔════╝╚══██╔══╝██║██║     ██║
+  ██║  ██║██║███████╗   ██║   ██║██║     ██║
+  ██║  ██║██║╚════██║   ██║   ██║██║     ██║
+  ██████╔╝██║███████║   ██║   ██║███████╗███████╗
+  ╚═════╝ ╚═╝╚══════╝   ╚═╝   ╚═╝╚══════╝╚══════╝
+              A N Y T H I N G
+     Distill any model into one you own
+</pre></div>
+
+<p align="center"><strong>generate → distill → judge → benchmark · white-box KD + black-box seqKD · any teacher (Claude / GPT / HF / Ollama) · LLM-as-judge report card · runs on a MacBook</strong></p>
+
+<p align="center">
+  <a href="https://github.com/AIAnytime/distillanything/actions/workflows/ci.yml"><img src="https://github.com/AIAnytime/distillanything/actions/workflows/ci.yml/badge.svg" alt="CI"></a>
+  <a href="https://www.python.org/"><img src="https://img.shields.io/badge/python-3.10%2B-blue" alt="Python 3.10+"></a>
+  <a href="LICENSE"><img src="https://img.shields.io/badge/license-Apache--2.0-green" alt="License: Apache 2.0"></a>
+  <a href="#development"><img src="https://img.shields.io/badge/tests-31%20passing%20offline-brightgreen" alt="Tests"></a>
+  <a href="#whats-real-today-vs-the-vision"><img src="https://img.shields.io/badge/status-alpha-orange" alt="Status: alpha"></a>
+  <a href="https://github.com/astral-sh/ruff"><img src="https://img.shields.io/badge/code%20style-ruff-261230" alt="Code style: ruff"></a>
+</p>
+
+<p align="center">
+  <a href="#get-started-60-seconds">Install</a> ·
+  <a href="#proof">Proof</a> ·
+  <a href="#the-report-card">Report card</a> ·
+  <a href="#teachers">Teachers</a> ·
+  <a href="#compared-to">Compared to</a> ·
+  <a href="#roadmap">Roadmap</a>
+</p>
+
+---
+
+Big models know things. Small models ship. Distill Anything covers the **whole distillation lifecycle** — a teacher generates your dataset, a student trains on its logits or its words, a judge scores the result blind, and a benchmark prices it — with one YAML recipe schema that scales from a 16GB MacBook to a GPU cluster.
+
+<p align="center">
+  <img src="https://raw.githubusercontent.com/AIAnytime/distillanything/main/media/demo.gif" alt="distill smoke running end-to-end on a MacBook" width="820">
+  <br/><sub>Live: <code>distill smoke</code> — full logit-KD pipeline, offline, &lt;60s on a MacBook. loss 2.24 → 2.08, PASS.</sub>
+</p>
+
+<p align="center">
+  <img src="https://raw.githubusercontent.com/AIAnytime/distillanything/main/media/architecture.png" alt="Distill Anything architecture: seed prompts through teachers, synthetic dataset pipeline, distillation engine, evaluation, benchmark" width="100%">
+  <br/><sub>The full picture. What's shipped vs planned: <a href="#whats-real-today-vs-the-vision">status table</a>.</sub>
+</p>
+
+## What it does
+
+- **Any teacher** — `claude`, `openai:gpt-4o-mini`, `ollama:llama3.2`, or any local `hf:` model, selected by one spec string
+- **Synthetic data** — `distill generate` turns seed prompts into a deduped instruction dataset, with self-instruct-style prompt expansion (`--expand 5`)
+- **Quality gating** — an LLM judge scores every record 1–10 before training (`--judge claude --min-score 7`)
+- **White-box logit KD** — forward KL (Hinton), reverse KL (MiniLLM-style), generalized JSD, temperature scaling, and top-k truncation so 50k-vocab KD fits in laptop memory
+- **Black-box seqKD** — no logits needed; fine-tune on what an API teacher wrote
+- **LLM-as-judge eval** — blind A/B, position-swapped to kill judge bias → win/tie/lose and one headline number: *quality retention*
+- **Report card** — `distill report` writes a shareable `REPORT.md`: quality vs the teacher, p50/p95 latency, tokens/s, memory, $ per 1K tokens
+- **Hardware-aware** — CUDA (bf16) → Apple Silicon MPS → CPU, detected automatically
+
+## How it works (30 seconds)
+
+```
+ Seed prompts (.txt / .jsonl)
+      │
+      ▼
+ Teacher ──────────── claude · openai:<m> · ollama:<m> · hf:<repo>
+      │  generate + expand + dedup + judge-score
+      ▼
+ Curated dataset (JSONL)
+      │
+      ▼
+ Distillation engine
+      ├─ logit KD    teacher logits → KL/JSD on every response token   (hf: teachers)
+      └─ seqKD       teacher text   → cross-entropy fine-tune          (any teacher)
+      │
+      ▼
+ Judge (blind A/B, position-swapped)  →  win / tie / lose
+      │
+      ▼
+ REPORT.md — "student keeps X% of teacher quality at 1/Nth the size and cost"
+```
+
+One `DistillConfig` YAML describes the whole run; the same recipe format works for a 135M student on a MacBook and a 7B student on an A100.
+
+## Get started (60 seconds)
+
+```bash
+# 1 — Install
+git clone https://github.com/AIAnytime/distillanything && cd distillanything
+pip install -e .                    # core (torch, transformers)
+pip install -e ".[anthropic]"       # optional: Claude as teacher/judge
+pip install -e ".[openai]"          # optional: OpenAI / vLLM / Ollama teachers
+
+# 2 — Verify your machine (tiny random models, zero downloads, <1 min)
+distill smoke
+
+# 3 — First real distillation (SmolLM2-360M → 135M, ~1GB download, runs on MPS)
+distill generate examples/data/seed_prompts.txt \
+  --teacher hf:HuggingFaceTB/SmolLM2-360M-Instruct --out data/train.jsonl --expand 5
+distill train recipes/mac-small.yaml
+distill report runs/mac-small --dataset data/train.jsonl \
+  --teacher hf:HuggingFaceTB/SmolLM2-360M-Instruct --judge hf:HuggingFaceTB/SmolLM2-360M-Instruct
+```
+
+Prefer Claude as the teacher? `export ANTHROPIC_API_KEY=...` then swap `--teacher claude` and `distill train recipes/claude-blackbox.yaml`.
+
+## Proof
+
+We publish only what we've measured. Every number below reproduces on a stock Apple-Silicon MacBook with the commands shown — no cluster required.
+
+**End-to-end pipeline, offline, under a minute** (`distill smoke`):
+
+```
+Training mode=logit device=mps steps=30 loss=forward_kl(T=2.0, alpha=0.5)
+  step 10/30  loss 2.2371   →   step 30/30  loss 2.0767
+PASS: loss 2.237 -> 2.077 over 30 steps
+```
+
+**Benchmark output** (`distill benchmark sshleifer/tiny-gpt2 --n-runs 3 --cost-per-hour 1.20`, measured on an M-series MacBook):
+
+| Metric | Value |
+|---|---:|
+| latency_p50_s | 0.054 |
+| latency_p95_s | 0.079 |
+| tokens_per_s | 262.4 |
+| memory_mb | 0.4 |
+| cost_per_1k_tokens_usd | 0.001271 |
+
+**A real distillation, end to end on one MacBook** — the exact Quickstart above (SmolLM2-360M teaches 135M on 60 records it generated itself, ~15 min total, all local):
+
+| | Student (135M, distilled) | Teacher (360M) |
+|---|---:|---:|
+| Parameters | 134.5M | 361.8M |
+| Tokens / s (MPS) | **70.5** | 56.7 |
+| Latency p50 | **0.40s** | 0.76s |
+| Disk | **538MB** | 1447MB |
+| $ / 1K tokens @ $1.20/hr | **$0.0047** | $0.0059 |
+
+- **Quality retention 87.5%** on 24 held-out prompts (1 win / 20 ties / 3 losses), judged blind with position swap by the teacher itself. A 360M judge is coarse — most pairs tie; point `--judge claude` at it for sharper discrimination.
+- Token-level teacher agreement **81%**, perplexity 2.17 (3-example eval split — demo scale).
+- The full unedited artifact: [examples/sample-report.md](examples/sample-report.md), generated by `distill report`.
+
+This is the quickstart, not a leaderboard claim — 60 records and 8 optimizer steps prove the loop runs on a laptop, not the ceiling of the method.
+
+**Honest-by-construction evals:** the judge sees answers blind and judges each pair twice with positions swapped — a judge that always prefers "Answer A" nets out to a tie (that exact adversarial case is in the test suite). Disagreements and unparseable verdicts count as ties, never wins. 31 tests run fully offline against tiny random models.
+
+## The report card
+
+The question that matters is *"did the student keep the teacher's quality at a fraction of the cost?"* — so every run can end with a one-page answer:
+
+```bash
+distill report runs/mac-small \
+  --dataset data/train.jsonl \
+  --teacher hf:HuggingFaceTB/SmolLM2-360M-Instruct \
+  --judge claude --n 32 --cost-per-hour 1.20
+```
+
+`REPORT.md` leads with **quality retention** (how often the student matches or beats the reference), then a side-by-side efficiency table (params, tokens/s, p50/p95, memory, $/1K tokens — "3.0x smaller and 3.0x faster"), sample outputs, and the training metrics. `report.json` sits next to it for machines.
+
+## Teachers
+
+One string selects any knowledge source — as teacher *or* as judge:
+
+| Spec | Backend | KD mode |
+|---|---|---|
+| `hf:HuggingFaceTB/SmolLM2-360M-Instruct` | local Hugging Face model | white-box (logits) |
+| `claude` / `claude:claude-opus-4-8` | Anthropic API | black-box (seqKD) |
+| `openai:gpt-4o-mini` | OpenAI API (or any compatible endpoint) | black-box |
+| `ollama:llama3.2` | local Ollama server | black-box |
+
+**One rule to know:** logit KD requires student and teacher to share a tokenizer (same model family). Across tokenizers, use `mode: seqkd` — cross-tokenizer logit alignment (ULD) is on the roadmap.
+
+## Recipes
+
+Everything is a YAML recipe (`distill init` writes a starter):
+
+```yaml
+mode: logit                # or seqkd
+teacher: { spec: hf:HuggingFaceTB/SmolLM2-360M-Instruct }
+student: { model: HuggingFaceTB/SmolLM2-135M-Instruct }
+data:    { path: data/train.jsonl, max_seq_len: 512 }
+loss:    { kind: forward_kl, temperature: 2.0, alpha: 0.5, top_k: 256 }
+train:   { output_dir: runs/out, lr: 1.0e-4, epochs: 2, batch_size: 2, grad_accum: 8 }
+```
+
+Or skip YAML entirely:
+
+```python
+from distillanything import Student
+
+student = Student("HuggingFaceTB/SmolLM2-135M-Instruct")
+student.learn(teacher="claude", dataset="data/prompts_only.jsonl")   # generates missing responses, then trains
+print(student.generate("Explain what a database index is."))
+print(student.benchmark())
+```
+
+## When to use · When to skip
+
+**Great fit if you…**
+- want a specialized model that's 10–100x smaller than the API model you're calling today
+- have (or can seed) a few hundred domain prompts and an API key or a local teacher model
+- work on a laptop — everything here is sized to run and iterate on 16GB of RAM
+
+**Skip it if you…**
+- need RAG or prompt engineering, not a trained model — distillation is for when the task is stable and volume is high
+- need cross-tokenizer logit KD, multimodal, or distributed training *today* (roadmap, not shipped)
+- expect a hosted platform — this is a framework; you bring the compute
+
+<a id="whats-real-today-vs-the-vision"></a>
+<details>
+<summary><b>What's real today vs the vision</b></summary>
+
+The architecture diagram is the north star, not the changelog. Every box, mapped honestly:
+
+**Legend:** ✅ shipped &nbsp;·&nbsp; 🚧 partial &nbsp;·&nbsp; 🗺️ planned, not yet built
+
+| Diagram section | Box | Status | Notes |
+|---|---|:---:|---|
+| Seed Prompts | Text / Docs, Custom | ✅ | `.txt` (one prompt/line) and `.jsonl` (`prompt`/`messages`/`text`) |
+| Seed Prompts | Code | 🚧 | No special handling — treated as plain text, works but untuned |
+| Teacher | Claude API, GPT API, Hugging Face (Local), Ollama (Local) | ✅ | `hf:` / `claude` / `openai:` / `ollama:` teacher specs |
+| Synthetic Dataset Pipeline | Generate, Deduplication, Curated Dataset | ✅ | `distill generate`, normalized-content dedup |
+| Synthetic Dataset Pipeline | Filtering | 🚧 | Empty/too-short response filter only — no content or toxicity filters yet |
+| Synthetic Dataset Pipeline | Quality Scoring | ✅ | LLM-judge 1-10 scoring: `distill generate --judge claude --min-score 7` |
+| Distillation Engine | Forward KL, Reverse KL, JSD, Top-k/Temp | ✅ | White-box logit KD, all three divergences + top-k truncation |
+| Distillation Engine | Response Supervision, Instruction Tuning | ✅ | Black-box seqKD (fine-tune on teacher-generated text) |
+| Distillation Engine | Preference (Optional) | 🗺️ | No DPO/preference-based distillation yet |
+| Evaluation | Perplexity | ✅ | |
+| Evaluation | Teacher Agreement | ✅ | Win/tie/lose via blind, position-swapped LLM judge (`distill report --judge`) + token-level agreement |
+| Evaluation | Accuracy/EM, BLEU/ROUGE/BERTScore, Safety/Bias Checks | 🗺️ | No task-benchmark or safety-eval harness yet |
+| Benchmark | Tokens/s, Memory, Model Size | ✅ | `distill benchmark` |
+| Benchmark | Latency | ✅ | p50/p95 over repeated runs (`--n-runs`) |
+| Benchmark | Cost / 1K Tokens | ✅ | From measured throughput × your `--cost-per-hour` |
+| Outputs | Model Weights, Tokenizer & Config, Training Logs, Benchmarks | ✅ | Saved to `output_dir` on every run |
+| Outputs | Evaluation Report | ✅ | `distill report` writes a shareable REPORT.md + report.json per run |
+| Core Capabilities | Reproducible Pipelines | ✅ | Seeded runs + full config snapshot saved alongside the checkpoint |
+| Core Capabilities | Multi-Teacher Support, Multi-Modal, Distributed Training, Quantization/Export, Experiment Tracking | 🗺️ | One teacher/one device per run today; text-only; no W&B/MLflow hooks |
+| Integrations | Hugging Face | ✅ | Models, tokenizers, chat templates |
+| Integrations | Weights & Biases, MLflow, S3/GCS/Azure Blob, Docker/Kubernetes | 🗺️ | Not integrated yet |
+| How you interact | Python SDK, CLI | ✅ | `Student().learn(...)` and `distill ...` |
+| How you interact | REST API, Web UI | 🗺️ | CLI/SDK only for now |
+
+</details>
+
+<details>
+<summary><b>What's inside</b></summary>
+
+- **`teachers/`** — one `Teacher` abstraction; local HF models expose logits (white-box), API teachers expose text (black-box). Registry resolves spec strings.
+- **`losses/kd.py`** — masked, temperature-scaled (T² gradient correction) forward KL / reverse KL / generalized JSD, with top-k support truncation and automatic vocab-padding alignment.
+- **`data/`** — JSONL/txt loading, chat-template rendering, prompt-masked tokenization, normalized dedup, teacher-driven generation with prompt expansion.
+- **`train/trainer.py`** — hand-rolled loop (no HF Trainer): grad accumulation, cosine LR + warmup, grad clipping, KD+CE mixing, MPS-safe autocast policy.
+- **`eval/judge.py`** — pairwise blind judging with position-swap debiasing; absolute 1–10 scoring for dataset gating.
+- **`eval/benchmark.py`** — p50/p95 latency, throughput, peak memory, $/1K tokens.
+- **`eval/report.py`** — REPORT.md + report.json builder.
+- **`testing.py`** — tiny random Llama models + in-memory char tokenizer, so the whole suite runs offline.
+
+</details>
+
+## Compared to
+
+Excellent distillation *trainers* exist — the gap this project fills is the **lifecycle around the training loop**:
+
+|  | Primary focus | Data generation from API teachers | Blind LLM-judge → report card | Sized for laptops |
+|---|---|:---:|:---:|:---:|
+| **Distill Anything** | Full lifecycle: generate → distill → judge → benchmark | ✅ | ✅ | ✅ |
+| [TRL](https://github.com/huggingface/trl) (GKD) | RL/KD training methods in the HF ecosystem | — | — | GPU-oriented |
+| [DistillKit](https://github.com/arcee-ai/DistillKit) | KD training techniques | — | — | GPU-oriented |
+| [EasyDistill](https://github.com/modelscope/easydistill) | KD training + data synthesis toolkit | ✅ | — | GPU-oriented |
+
+If you already have a curated dataset and a GPU and just want a training loop, TRL's GKD trainer is great — and our loss implementations follow the same literature (Hinton KD, MiniLLM, DistiLLM).
+
+## Roadmap
+
+Beyond closing the 🗺️ gaps in the status table:
+
+- [ ] **LoRA/QLoRA students** — distill into 1–3B students on 16GB of RAM *(next up)*
+- [ ] Hidden-state / feature KD with learned projectors
+- [ ] Cross-tokenizer logit distillation (ULD)
+- [ ] Multi-teacher voting and ensembling
+- [ ] VLM, embedding, and reranker distillation
+- [ ] Eval harness integration (lm-eval-harness) and regression tracking
+
+Everything in this repo is and stays Apache-2.0. A hosted layer (GPU orchestration, dashboards, model registry) may come later — the framework is the product, not the funnel.
+
+## Development
+
+```bash
+uv venv && uv pip install -e ".[dev]"
+pytest            # 31 tests, fully offline (tiny random models, fake judges)
+distill smoke     # end-to-end pipeline check on your hardware
+```
+
+PRs welcome — the test suite needs no GPU, no API keys, and no downloads.
+
+## License
+
+Apache 2.0 — see [LICENSE](LICENSE).