Real research papers, rewritten in plain words by small open models we trained. Live at sciencemadereadable.com
See the codehttps://sciencemadereadable.com: search open research papers and read them rewritten in plain words, next to the original, with the paper's own figures and tables.
The rewrite is a translation, not a summary: the same paper, in the authors' voice, with the same information, uncertainty, headings and order, written so that a curious 14-year-old can follow it. It is done by a small model we trained (Qwen3.5 9B, fine-tuned on about 4,400 papers rewritten by large models), which runs on a single GPU.
training/figures/blog/x_article.md, with its figures in training/figures/blog/out/paper_corpus/ open-licence papers (Europe PMC): harvest, split into sections, large-model rewrites
glossary/ reference entries for each paper (offline Wikipedia and Wiktionary), given to the model
rewrite_benchmark/ the rewrite recipe (prompts) and the judge that scores rewrites
model_baselines/ other models' rewrites and scores on the same test papers
training/ training the small models on the corpus; speed and quality measurements
v3/ the next training set (in progress): every fact from the paper, the glossary,
or what a 14-year-old knows; ⟦markers⟧ for anything else
app/ the website, the job queue, the GPU worker (app/README.md)
notebooks/ tutorials: how the rewrite recipe was built and tested
explainers/, article-tokens/, pine-nut-visuals/ side experiments
The prompts are the same everywhere: the app's TypeScript programs send byte for byte what the
Python training and benchmark code sent (app/tools/check_prompts.ts), so the model in production
reads exactly what it was trained and scored on.
Python parts (from this folder):
uv sync
.venv/bin/python app/server.py # queue, paper service and the page: http://127.0.0.1:8795
The page and the worker (Node 22+). They use functai's TypeScript library, which is not on npm yet: clone it next to this folder first.
git clone https://github.com/MaximeRivest/functai ../functai
(cd ../functai/ts && npm install)
cd app && npm install && node tools/build.mjs
npm run worker # takes jobs, runs the model through vLLM (app/worker/)
Serving the model: any OpenAI-compatible server, for example
vllm serve maximerivest/qwen3.5-9b-paper-rewriter --dtype bfloat16 --max-model-len 65536 --speculative-config '{"method":"mtp","num_speculative_tokens":2}' --override-generation-config '{"temperature":0.0}' --default-chat-template-kwargs '{"enable_thinking":false}'.
app/rent/h100_up.sh rents one H100 on Nebius and does all of this.
Data and models live next to the code on the working machine and are left out (.gitignore):
paper copies, the large-model rewrites, the offline Wikipedia copy (46 GB,
glossary/data/offline/download.sh), training data and checkpoints (about 200 GB), and run outputs.
The trained models are on Hugging Face (above).
training/v3/ builds the next training set: every fact in an answer must come from the paper, the
glossary given with it, or what a curious 14-year-old knows; anything else becomes a ⟦marker⟧ to be
explained afterwards, so the small model writes instead of recalling. Design: training/dataset_v3.md.
A 50-paper pilot (training/v3/pilot50.py) finished on 2026-10-05: in a blind comparison the judge
still preferred the current answers (159 to 40, 41 ties), so the recipe needs more work before training.
A research project that went live on 2026-10-05. The rewrites read well and keep the papers' structure, but they can still contain factual errors: always check the original, shown alongside.
The site is free, with no ads and no accounts; a rented GPU writes the rewrites, about $108 a day. Readers keep it running: sciencemadereadable.com/?support (card, Apple Pay or Google Pay through Stripe, or GitHub Sponsors). Organizations can sponsor a whole day and be named on the site that day. This is support for an independent project, not a charity donation.
Code and notes: Apache 2.0. The papers shown and rewritten are open-access articles under CC BY; their rewrites credit the original authors and link to the paper.
Real research papers, rewritten in plain words by small open models we trained. Live at sciencemadereadable.com
See the codehttps://sciencemadereadable.com: search open research papers and read them rewritten in plain words, next to the original, with the paper's own figures and tables.
The rewrite is a translation, not a summary: the same paper, in the authors' voice, with the same information, uncertainty, headings and order, written so that a curious 14-year-old can follow it. It is done by a small model we trained (Qwen3.5 9B, fine-tuned on about 4,400 papers rewritten by large models), which runs on a single GPU.
training/figures/blog/x_article.md, with its figures in training/figures/blog/out/paper_corpus/ open-licence papers (Europe PMC): harvest, split into sections, large-model rewrites
glossary/ reference entries for each paper (offline Wikipedia and Wiktionary), given to the model
rewrite_benchmark/ the rewrite recipe (prompts) and the judge that scores rewrites
model_baselines/ other models' rewrites and scores on the same test papers
training/ training the small models on the corpus; speed and quality measurements
v3/ the next training set (in progress): every fact from the paper, the glossary,
or what a 14-year-old knows; ⟦markers⟧ for anything else
app/ the website, the job queue, the GPU worker (app/README.md)
notebooks/ tutorials: how the rewrite recipe was built and tested
explainers/, article-tokens/, pine-nut-visuals/ side experiments
The prompts are the same everywhere: the app's TypeScript programs send byte for byte what the
Python training and benchmark code sent (app/tools/check_prompts.ts), so the model in production
reads exactly what it was trained and scored on.
Python parts (from this folder):
uv sync
.venv/bin/python app/server.py # queue, paper service and the page: http://127.0.0.1:8795
The page and the worker (Node 22+). They use functai's TypeScript library, which is not on npm yet: clone it next to this folder first.
git clone https://github.com/MaximeRivest/functai ../functai
(cd ../functai/ts && npm install)
cd app && npm install && node tools/build.mjs
npm run worker # takes jobs, runs the model through vLLM (app/worker/)
Serving the model: any OpenAI-compatible server, for example
vllm serve maximerivest/qwen3.5-9b-paper-rewriter --dtype bfloat16 --max-model-len 65536 --speculative-config '{"method":"mtp","num_speculative_tokens":2}' --override-generation-config '{"temperature":0.0}' --default-chat-template-kwargs '{"enable_thinking":false}'.
app/rent/h100_up.sh rents one H100 on Nebius and does all of this.
Data and models live next to the code on the working machine and are left out (.gitignore):
paper copies, the large-model rewrites, the offline Wikipedia copy (46 GB,
glossary/data/offline/download.sh), training data and checkpoints (about 200 GB), and run outputs.
The trained models are on Hugging Face (above).
training/v3/ builds the next training set: every fact in an answer must come from the paper, the
glossary given with it, or what a curious 14-year-old knows; anything else becomes a ⟦marker⟧ to be
explained afterwards, so the small model writes instead of recalling. Design: training/dataset_v3.md.
A 50-paper pilot (training/v3/pilot50.py) finished on 2026-10-05: in a blind comparison the judge
still preferred the current answers (159 to 40, 41 ties), so the recipe needs more work before training.
A research project that went live on 2026-10-05. The rewrites read well and keep the papers' structure, but they can still contain factual errors: always check the original, shown alongside.
The site is free, with no ads and no accounts; a rented GPU writes the rewrites, about $108 a day. Readers keep it running: sciencemadereadable.com/?support (card, Apple Pay or Google Pay through Stripe, or GitHub Sponsors). Organizations can sponsor a whole day and be named on the site that day. This is support for an independent project, not a charity donation.
Code and notes: Apache 2.0. The papers shown and rewritten are open-access articles under CC BY; their rewrites credit the original authors and link to the paper.