Datasetpublic

Akash-Sakala/transcript-formatter-curriculum

Transcript Formatter Curriculum (L0–L5 + RL) Training data behind Akash-Sakala/gpt-oss-120b-transcript-formatter-lora: a layered curriculum that turns raw speech-to-text transcripts into clean, formatted transcripts. Every row is input (raw transcript) → output (formatted), across 21 categories spanning punctuation, casing, fillers, disfluencies, homophones, ITN, proper nouns, URLs/emails, and layout. Subsets (use the Data Viewer dropdown) Each curriculum level is… See the full description on the dataset page: https://huggingface.co/datasets/Akash-Sakala/transcript-formatter-curriculum.

sourceHugging Faceapache-2.0updated 3mo agoView on Hugging Face

0likes10downloads

Dataset Card

Transcript Formatter Curriculum (L0–L5 + RL)

Training data behind `Akash-Sakala/gpt-oss-120b-transcript-formatter-lora`: a layered curriculum that turns raw speech-to-text transcripts into clean, formatted transcripts. Every row is input (raw transcript) → output (formatted), across 21 categories spanning punctuation, casing, fillers, disfluencies, homophones, ITN, proper nouns, URLs/emails, and layout.

Subsets (use the Data Viewer dropdown)

Each curriculum level is a separate config, with a train split (*_train.csv) and a validation split (*_val.csv). L0 is the base set with a train split only.

Config	Adds	Train	Validation
L0	basics (punctuation, capitals, passthrough)	`sft_l0/L0.csv`	—
L1	surface edits (filler, homophone, ITN, comma, date)	`sft_l1/L1_train.csv`	`sft_l1/L1_val.csv`
L2	disfluency (stutter, false-start, backtrack)	`sft_l2/L2_train.csv`	`sft_l2/L2_val.csv`
L3	artifacts (proper nouns, URLs/emails)	`sft_l3/L3_train.csv`	`sft_l3/L3_val.csv`
L4	layout (emails, lists, symbols)	`sft_l4/L4_train.csv`	`sft_l4/L4_val.csv`
L5	corrective SFT (over-format fix, gold cleanup)	`sft_l5/L5_train.csv`	`sft_l5/L5_val.csv`
RL	RL-alignment prompts (archetype-verifiable)	`sft_rl/rl_align_train.csv`	`sft_rl/rl_align_val.csv`

Each *_train split bakes in replay of all prior-layer categories (continued SFT forgets un-replayed categories), so a single config is self-contained.

Schema

Curriculum configs (L0–L5): id, category, subtype, domain, dialect, register, length_bucket, input, output

RL config: id, archetype, reward_type, category, length_bucket, input, expected

—input — raw transcript (Whisper-style output).
—output / expected — the formatted target.
—category — one of 21 (e.g. sent_end_punct, filler_removal, stutter_removal, proper_noun_fix, email_block, … and passthrough, the negative-trap boundary).
—length_bucket — small / medium / large / xl. Multi-dialect (en-US/GB/AU/IN/CA/IE/ZA).
—passthrough rows are negative-traps: output == input (the model must NOT edit).

Load with 🤗 datasets

python

from datasets import load_dataset

# one level (config) with its splits
l4 = load_dataset("Akash-Sakala/transcript-formatter-curriculum", "L4")
print(l4)                 # DatasetDict({ train: ..., validation: ... })
print(l4["train"][0])

# L0 has only a train split
l0 = load_dataset("Akash-Sakala/transcript-formatter-curriculum", "L0", split="train")

License Apache-2.0 (matches the base openai/gpt-oss-120b).