Open release

Here are the receipts.

The dataset, the scripts that built it, the trained weights, and the result of the run. Where every row came from and how it was checked. Four commands you can run yourself to confirm the counts. And the training result, which was a regression, reported with the reason rather than buried.

The artifacts

Six links, all live.

Dataset on two hosts, weights on two hosts, the pipeline that built the dataset, and a running interface. Each card says what is inside, so you know what you are getting before you click.

What is inside

17,000 rows, and where each one came from.

Six columns, all populated, zero missing values anywhere. Two of them do the real work. source names the corpus a row came from. verify_level says how that row was checked, or admits that it was not. Adaption ingested 17,586 rows from this corpus for the training run.

Check it yourself

Do not take my word for it. Run these.

Four calls. No account, no key, no login. Each one settles a number I have put on this page, and the output you should get back is printed next to it. If I have something wrong, one of these catches me.

1. 01

   The dataset has 17,000 rows and 6 columns.

   Hugging Face computes this from the parquet file itself. It is not a number I typed into a card.

   runnable

   the command

   curl -s "https://datasets-server.huggingface.co/size\ ?dataset=manifesta/verified-math-code-17k" \ | python -c "import sys,json; s=json.load(sys.stdin)['size']['configs'][0]; \ print(s['num_rows'], s['num_columns'])"

   what it returns

   17000 6

   If the published file ever stops matching the row count I state, this command says so before I do.

2. 02

   13,260 of 17,000 rows carry verification, and 0 values are missing.

   The same service counts every distinct value in every string column, source and verify_level included.

   runnable

   the command

   curl -s "https://datasets-server.huggingface.co/statistics\ ?dataset=manifesta/verified-math-code-17k&config=default&split=train"

   what it returns

   verify_level independent 9104 · executed 4156 · "-" 3740 source OpenMathReasoning 4380 · OpenCodeInstruct 3770 OpenMathInstruct-2 3248 · Tulu-3 3108 OpenR1-Math 1476 · Dolly-15k 632 code_contests 386 nan_count 0 on all six columns

   9,104 plus 4,156 is 13,260 of 17,000, which is 78.0 percent. The other 3,740 are general rows that carry no verification. They are marked in the data, not hidden in it.

3. 03

   The weights exist and download.

   An empty weights repo looks identical to a full one from the outside. Ask the CDN for the file size.

   runnable

   the command

   curl -sIL https://huggingface.co/manifesta\ /adaption_verified_math_code_instruct\ /resolve/main/adapter_model.safetensors \ | grep -i "^HTTP\|^content-length"

   what it returns

   HTTP/1.1 302 Found HTTP/1.1 200 OK content-length: 43856576

   43,856,576 bytes of LoRA adapter. trainer_state.json sits beside it in the same repo, and that is where every training number below comes from.

4. 04

   The declared base model resolves.

   adapter_config.json names a base model. That name has to be a repo that exists, or nothing can load the adapter at all.

   runnable

   the command

   curl -s -o /dev/null -w "%{http_code}\n" \ https://huggingface.co/api/models/google/gemma-4-31B-it

   what it returns

   200

   adapter_config.json says base_model_name_or_path is google/gemma-4-31B-it. The model card says the same thing. The Hub returns 200. All three agree.

Rather click than type. The raw endpoints: size, statistics, adapter_config.json, and the base model.

The training result

The adapted model lost.

AutoScientist scored it against the base, head to head. Adapted 46. Base 54. It went backwards by 8 points and there is no reading of that which makes it a win. So here is what happened, from trainer_state.json inside the published weights.

Reproduce

The dataset is the output of a pipeline, and the pipeline ships.

Nine scripts, published in the dataset repo under build/. This is not a hand curated pile with a story attached to it. Run them in this order and you get the file.

All nine, the audit evidence, and a verifier that rechecks every number on this page against the live published files, are in one public repo: A1VARA5/verified-math-code-17k. verify.py is standard library only, so there is no install step and no account. It runs 23 checks, and the same 23 run on a daily schedule, which means if anything on this page stops being true the badge goes red without me touching it.