Training datasets · 100% local
Fine-tuning a model or building a RAG on client documents means building a dataset, and today that means uploading the PDFs to a service or wiring a Python pipeline designed for terabytes of web text. Sırdaş runs the whole chain in your browser or your terminal: convert, redact, deduplicate, split by document and export in the format your trainer expects.
PDF converters end at Markdown and corpus curators start at JSONL. The bridge between them, with provenance, per-document splits and a redaction trail, is what we built.
Each one has literature documenting it. Each one has a step in the pipeline that fixes it.
Over 1% of what a model trained on public corpora emits is verbatim training text. A folder of contracts on the same template is the extreme case.
Deduplication cuts memorized text tenfold and needs fewer training steps for the same quality.
What we do
MinHash + LSH over 5-grams, 128 signatures, threshold 0.8, final check with exact Jaccard.
If a fragment of the evaluation set is in the training set, the score measures memory, not learning. Llama 3 published benchmarks contaminated up to 98%.
GPT-3 set the method: 13-gram overlap, discarding n-grams that appear in more than 10 documents.
What we do
sirdas contamination, 13-grams by default, and the literal matching fragment in the report so the finding can be checked.
Chunk first and split later, and two fragments of the same contract land one in train and one in test. The metric improves, and lies.
Kapoor and Narayanan count 329 papers affected by leakage across 17 fields.
What we do
The split is always by document, stable (same seed, same split), and adding a document never moves the ones already there.
Names, phone numbers and emails that appeared in a single corpus document were extracted from GPT-2. Larger models are more vulnerable.
Nobody detects generic identifiers reliably: StarCoder's human annotators reached 53%. Our redaction is rule-based and can fail too.
What we do
Redaction on by default, with a count of replacements and a warning if you turn it off. Reversible with the mapping, which stays with you.
Every trainer expects a different schema and the error shows up late. Field names are verified against the official docs.
| Format | Who consumes it | Fields |
|---|---|---|
| text | Continued pretraining (TRL) | text |
| chat | TRL, Unsloth, OpenAI fine-tuning | messages[{role, content}] |
| prompt-completion | TRL | prompt, completion |
| alpaca | Axolotl, LLaMA-Factory, Unsloth | instruction, input, output |
| sharegpt | Chat trainers | conversations[{from, value}] |
| langchain | Chains and vector stores | page_content, metadata |
| llamaindex | TextNode with stable id | id_, text, metadata |
| embeddings | Pinecone, Qdrant, pgvector, Chroma | id, text, metadata |
| train / validation / test | Anything, with a Hugging Face card (--hf) | three files + README.md |
| RLVR (sirdas rl) | GRPOTrainer and verifiable rewards | prompt, ground_truth |
The web tool is for a folder. For thousands of documents the CLI does the same on your machine, with --json output for scripts.
See the CLIConvert, redact, dedupe and split. With --hf you get train/validation/test and the Hugging Face card.
npx sirdas dataset docs/*.pdf --format chat --split --hf -d dataset/ npx sirdas dataset contracts/*.docx --format alpaca --strip-boilerplate -d ds/
And why. Anything that requires sending your documents to a model contradicts the reason this product exists.
There is a clear dividing line: converters end at Markdown and curators start at JSONL.
| Tool | Local | Dedup | Contamination | PII | Exports for training |
|---|---|---|---|---|---|
| unstructured, Docling, marker, nougat | |||||
| LlamaParse | no, cloud | ||||
| datatrove (HF) | yes, Python | regex | |||
| dolma (AI2) | yes, Python + Rust | regex | |||
| NeMo Curator | yes, Python + CUDA | Presidio | |||
| text-dedup | yes, Python | ||||
| Sırdaş | browser / Node | yes, reversible |
jsonlkit.com already does deduplication, PII masking and leak detection in the browser; what it lacks is PDF ingestion. The techniques (MinHash, LSH, n-grams, group splits) have been published for years: what is new is that they run on your PDFs with nothing to install and the file never leaving your machine. Full research, with sources, in the document linked below.
Read the full research (with sources)The three articles that explain, with citations, what almost everyone gets wrong the first time.
From PDF to JSONL: extract the text, redact, drop duplicates, split by document and export in the format your trainer expects. All on your machine.
9 minDuplicates, contamination, leakage through chunking and personal data: four documented mistakes, how to fix each, and why the third one fools your metric.
11 minThe field names each trainer expects, verified against its official documentation, and the details that break a file without telling you why.
10 minOut comes a JSONL you can audit, with your data exactly where it was: on your machine.
Build a dataset now