sandesh-bench / data /curated /dev /README.md
ashank-crazy's picture
Upload folder using huggingface_hub
159ac39 verified
|
Raw History Blame Contribute Delete
2.35 kB
metadata
tags:
  - curatorkit
configs:
  - config_name: sft_messages
    default: true
    description: >-
      curatorkit sft_messages export, sha256
      9d02a1d957ee89a4fd560e3b33e8ece56288500ade0b0dd40b3cb845bd38523e
    data_files:
      - split: train
        path: sft_messages.jsonl

Dataset Card: sandesh_dev

Summary

  • Run timestamp: 2026-11-15T10:47:02.546817Z
  • Wall clock time: 3.6s
  • Pipeline config hash: 8b0d2e9f1282abec

Source Documents

  • Unknown

Pipeline Stages

Step Input Passed Recovered Rejected Forward
JSONLReader 140 140 — 0 140
SchemaGate 140 140 0 0 140
ExactDeduplicator 140 140 — — 140
MinHashDeduplicator 140 140 — — 140
TextCleaner 140 140 — — 140
MessagesExporter — 140 — — 140

Gate Pass Rates and Rejection Breakdown

  • JSONLReader: 140/0 passed (N/A), 0 rejected
  • SchemaGate: 140/140 passed (100.0%), 0 rejected

Rejection reasons:

  • None

Deduplication

  • near_duplicates_removed: 0
  • surviving_count: 140

Known Limitations

  • MinHash threshold used: 0.95 — this is a starting point, not a universal answer. Adjust per domain (legal, medical, and other domain-specific corpora may need a lower threshold to avoid losing structurally similar but semantically valid examples).
  • Table extraction requires extract_tables: true on the PDF reader and MinerU with GPU support. Disabled by default.
  • Token counting uses a whitespace tokenizer by default. For exact LLM context window enforcement, enable use_tiktoken: true in the SchemaGate config.
  • Both within-run (MinHash/Diversity) and cross-run (EmbeddingDeduplicator) deduplication are available. Cross-run dedup persists an embedding index across runs; use embedding_reset_index: true to start fresh.

Reproduction

pip install curatorkit
curatorkit run pipeline.yaml --output-dir ./out/

Tool Versions

  • curatorkit: 0.1.28
  • python: 3.12.3

Configs

Load one with datasets.load_dataset("<this folder or Hub repo>", "<config>").

Config Splits (rows) Columns
sft_messages train (140) messages: list of {role, content}; role is system, user or assistant