orb-planner-data / README.md
marcosmaceo's picture
ORB planner data (0.1.0)
4217f33 verified
|
Raw History Blame Contribute Delete
3.89 kB
metadata
license: apache-2.0
pretty_name: ORB planner data
task_categories:
  - text-to-audio
language:
  - en
tags:
  - audio
  - sound-effects
  - synthetic
  - procedural-audio
  - orb
size_categories:
  - 100K<n<1M
configs:
  - config_name: grammar
    default: true
    data_files:
      - split: train
        path: data/grammar/train.jsonl.gz
      - split: test
        path: data/grammar/test.jsonl.gz
  - config_name: freetext
    data_files:
      - split: train
        path: data/freetext/train.jsonl.gz
      - split: validation
        path: data/freetext/validation.jsonl.gz

ORB planner data

Code · Planner · Demo

The 300,000 training rows behind the released ORB planner (v8, 47 roles, 13 textures), plus the held-out sets used to score it. Everything is procedurally generated text. No audio and no third-party content is included.

Config Split Rows Row shape
grammar train 150,000 {"caption", "score"}
grammar test 500 {"caption", "score"}
freetext train 150,000 {"caption", "spec"}
freetext validation 500 {"caption", "spec"}

v8 was warm-started from v7 and trained on grammar/train and freetext/train concatenated, in that order. It scores 500/500 exact on grammar/test.

Row shapes

grammar: a caption in ORB's exact caption grammar and the full renderer input (score). The planner is trained on the compact Spec derived from it (orb.grammar.spec_from_score).

{"caption": "5.00-second bright cue, vinyl bed, with fireworks @0.09 4.37s -10dB R1 #a8, wave crash @0.12 3.88s -10dB R1 #b1, ...",
 "score": {"duration": 5.0, "sample_rate": 48000, "mood": "bright", "textures": [...], "events": [...], ...}}

freetext: a plain-words prompt and the target Spec itself.

{"caption": "sound effect of a steel on steel on a mountain",
 "spec": {"duration": 3.0, "texture": "wind", "mood": "intense",
          "events": [{"role": "sword_clash", "t": 0.5, "dur": 1.5, "gain_db": -8.0, "pan": 0.0, "seed": 0}]}}

Loading

from datasets import load_dataset

grammar = load_dataset("marcosmaceo/orb-planner-data", "grammar", split="train")
freetext = load_dataset("marcosmaceo/orb-planner-data", "freetext", split="train")
print(freetext[0]["caption"], freetext[0]["spec"])

How it was made

From the repository, at tag v0.1.0:

python scripts/gen_scores.py --n 150000 --seed 2 --no-music --out data/scores_150k_v7.jsonl    # grammar/train
python scripts/gen_scores.py --n 500 --seed 54321 --no-music --out data/heldout_v7.jsonl       # grammar/test
python scripts/gen_freetext.py --n 150000 --seed 11 --out data/freetext_150k.jsonl             # freetext/train

The two train files rebuild byte for byte from these commands. freetext/validation is a separate draw from the same generator. gen_freetext.py never emits a prompt from the hand-written 209-prompt eval set (data/freetext_eval.jsonl in the repo), which is kept out of this dataset on purpose.

Intended use

Training and evaluating caption-to-Spec planners for the ORB engine. The grammar mapping is deterministic, so the grammar config also works as a clean sequence-copying benchmark.

Limitations

  • grammar captions are rigid by design; they are not natural language.
  • freetext prompts come from templates and word lists (orb.freetext), so they cover everyday phrasing narrowly. Real user prompts are messier.
  • No music, and no vocal sounds.

License

Apache-2.0. Generated by code in the ORB repository; contains no third-party material.