EXt1/comi-mixed-100k-jsonl

COMI mixed 100K: text-only JSONL

One plain UTF-8 .jsonl file per split, with one example per line: train.jsonl, eval.jsonl, test.jsonl, smoke_train.jsonl, smoke_eval.jsonl. No tokenizer, token-ID columns, Arrow files or model weights are distributed.

Split	Rows
train	100,000 (20,000 from each source)
eval	1,000 (200 from each source)
test	1,000 (200 from each source)
smoke_train	32
smoke_eval	10

Columns: input (context string), prompt (question/instruction string), answer (nonempty list of reference strings), source, id, and context_truncated (boolean). COMI training uses the first reference. Text and reference alternatives match the existing prepared mixture.

from datasets import load_dataset

# For a private repository, authenticate first with `hf auth login`.
ds = load_dataset("EXt1/comi-mixed-100k-jsonl", token=True)
train = ds["train"]
evaluation = ds["eval"]

# Or stream text without downloading the entire dataset at once:
stream = load_dataset("EXt1/comi-mixed-100k-jsonl", split="train", streaming=True, token=True)
row = next(iter(stream))
print(row.keys())

Choose your model/tokenizer in the training pipeline. This export performs no tokenization. The text was previously capped using Qwen/Qwen2-7B-Instruct's 32,768-token budget (context + prompt + longest reference including EOS). Long contexts were prefix-truncated; questions/answers were retained. This export does not recover the original untruncated contexts. Token counts can change with another tokenizer: retokenize and recheck limits before training.

This repository is a separate text dataset. The earlier runpod_terminal.py loads pretokenized Arrow from its original dataset; it does not automatically switch to this JSONL repository. Use load_dataset here and tokenize in your chosen training pipeline, or download JSONL files for the upstream COMI loader.

The training set follows COMI's described five-source recipe, but is not the paper's exact sampled rows. NQ uses the Tevatron DPR passage variant. NarrativeQA uses full stories before length capping. All supplied HotpotQA/2Wiki context is included before capping. MultiNews uses original paired document/summary files. Seed: 42. NQ/HotpotQA/2Wiki eval/test are distinct labeled dev/validation samples, not their official benchmark tests. Smoke rows intentionally overlap their parent splits. HR example datasets are excluded. Token-count metadata was removed; truncation provenance remains. No new tokenizer is run during export.

Sources and original terms

No new unified license is assigned to the source texts. Refer to each original dataset's card and terms; attribution and pinned revisions follow.

Source	Dataset card / original terms	Pinned revision
NaturalQuestions	Tevatron/wikipedia-nq	37a7d5bb286c76f1b406d76a7434e8eeb4385203
HotpotQA	hotpotqa/hotpot_qa	1908d6afbbead072334abe2965f91bd2709910ab
2WikiMQA	voidful/2WikiMultihopQA	16852fde9d85cba158cf7e6517e7a3f9415a28c0
NarrativeQA	deepmind/narrativeqa	2e643e7363944af1c33a652d1c87320d0871c4e4
MultiNews	alexfabbri/multi_news	38cb206959a2cca3aa49858914fba76258a3dcaf

manifest.json records per-file SHA256, byte counts, row counts, source balance, source provenance and inherited length policy. Local export checks original hashes/schema, split isolation and smoke membership before upload.