Datasets:

mkubaszek
/

aispiritlabs-fine-tune-workshop

source stringlengths 22 37	target stringlengths 7 12
the cat sat on the mat	cat sat
the dog ran very fast today	dog ran
a small bird flies high above	bird flies
the sun is bright in the sky	sun shines
my friend likes warm green tea	tea is warm
she reads a very good book	she reads
we walk slowly to the park	walk to park
a cup of cold fresh water	cold water
the moon rises high at night	moon rises
my brother plays a soft song	plays a song
the wind blows very softly today	wind blows
a small child draws happy faces	child draws
the rain falls down on red roofs	rain falls
our small team won the big game	team won
the river flows into the deep sea	river flows
the birds sing sweetly in the morning	birds sing

Fine-Tune Workshop Fixtures

Tiny hand-crafted JSONL fixtures used as training/eval data across the fine-tune-workshop repo. Every fixture is also published as a Parquet config for direct load_dataset(..., name="<slug>") consumption.

Configs

Config	Source lab	Summary
`hello-6pair`	`foundations/00_hello_fine_tune`	6 prompt/response pairs for the hello-world fine-tune lab.
`raw_explanations`	`foundations/01_data_foundations`	Raw explanation records used by the data-foundations lab.
`explanations_tokenization`	`foundations/02_tokenization`	Explanation corpus used to compare tokenization strategies.
`explanation_quality`	`foundations/04_metrics_and_comet`	Labeled explanations for the metrics + Comet foundations lab.
`concepts-pretrain-10`	`decoder/10_pretrain`	12 concept records with 8-field schema used by the decoder pretrain lab (beginner/intermediate).
`buckets-pretrain-10-adv`	`decoder/10_pretrain`	Difficulty-bucketed sentences for the decoder pretrain advanced tier's curriculum comparison.
`concepts-sft-12`	`decoder/12_sft`	8 multi-field concept records used by the decoder SFT lab.
`concept-bank-lora-13`	`decoder/13_lora`	8-concept bank for the decoder LoRA beginner + intermediate tiers.
`concepts-lora-13-adv`	`decoder/13_lora`	4-concept fixture for the decoder LoRA advanced tier's quantization frontier.
`concept-bank-judges-16-beg`	`decoder/16_evals_and_judges`	8-concept bank for the decoder evals-and-judges beginner tier.
`concept-bank-judges-16-int`	`decoder/16_evals_and_judges`	10-concept bank for the decoder evals-and-judges intermediate tier (adds qlora + overfitting).
`concepts-judges-16-adv`	`decoder/16_evals_and_judges`	4-concept fixture for the decoder evals-and-judges advanced tier (REINFORCE + KL capstone).
`mlm-8sentence`	`encoder/30_mlm_pretrain`	8-sentence corpus for the encoder MLM pretrain lab.
`sentiment-10pair`	`encoder/31_classification (shared with encoder/32_lora_encoder)`	10 balanced sentiment pairs. Shared across encoder/31_classification and encoder/32_lora_encoder.
`span-8triple`	`encoder/33_span_extraction`	8 sentence/start/end triples for the encoder span-extraction lab.
`abstractive-16pair`	`seq2seq/40_sft_abstractive (shared across seq2seq/40-48)`	16 source/target pairs. Shared across the seq2seq track (labs 40-48) and the 47_dpo synth-track handoff.

Usage

from datasets import load_dataset

# Pick a config by name:
ds = load_dataset("mkubaszek/aispiritlabs-fine-tune-workshop", name="hello-6pair", split="train")
ds = load_dataset("mkubaszek/aispiritlabs-fine-tune-workshop", name="raw_explanations", split="train")
ds = load_dataset("mkubaszek/aispiritlabs-fine-tune-workshop", name="explanations_tokenization", split="train")

Or set USE_LOCAL_DATASET=0 in the corresponding workshop lab and the loader pulls its config from this repo automatically.

Files

Each config lands under data/<slug>/:

data/<slug>/<slug>.parquet — the canonical load_dataset target.
data/<slug>/<slug>.jsonl — the learner-inspectable source fixture verbatim.

Files are named after the slug rather than by split convention; every config exposes one blob under HF's default train split. Workshop labs load the full blob and do their own train/val/test partitioning in Python.

Source

Fixtures are generated from the fine-tune-workshop repo. The .jsonl files are the source of truth; the .parquet files are derived via datasets.Dataset.from_json(...).to_parquet(...).

Downloads last month: 28