lca0503's picture
Preprocessed translation dataset with final prompts
f1b34c9
|
Raw History Blame Contribute Delete
5.88 kB
---
dataset_info:
config_name: source
features:
- name: id
dtype: string
- name: dataset
dtype: string
- name: source_text
dtype: string
- name: source_language
dtype: string
- name: target_language
dtype: string
- name: constraint_source
dtype: string
- name: target_length
dtype: int64
- name: rhyme
struct:
- name: finals
list: string
- name: id
dtype: int64
- name: boundary_mask
dtype: string
- name: required_word_boundaries
list: int64
- name: boundary_unit
dtype: string
- name: messages
list:
- name: content
dtype: string
- name: role
dtype: string
- name: prompt
dtype: string
- name: reference
struct:
- name: id
dtype: string
- name: scope
dtype: string
- name: text
dtype: string
- name: type
dtype: string
splits:
- name: validation
num_bytes: 14267180
num_examples: 4011
- name: test
num_bytes: 14272721
num_examples: 4006
download_size: 15765741
dataset_size: 28539901
configs:
- config_name: source
data_files:
- split: validation
path: source/validation-*
- split: test
path: source/test-*
license: cc-by-nc-sa-4.0
tags:
- lyrics
- controllable-generation
language:
- en
- zh
task_categories:
- translation
---
# ControllableLyricTranslation
English → Mandarin text-LLM inputs derived from
[LongshenOu/lyric-trans-en2zh-data](https://huggingface.co/datasets/LongshenOu/lyric-trans-en2zh-data),
released with [ControllableLyricTranslation](https://github.com/Sonata165/ControllableLyricTranslation).
This derivative packages the released `data_parallel` corpus by default, with
prompts and supplied controls. A custom corpus can be explicitly selected.
The upload pipeline keeps this repository private.
## Load
Log in with `hf auth login`, using an account with access to this repository.
```python
from datasets import load_dataset
data = load_dataset("lca0503/ControllableLyricTranslation", name="source", token=True)
row = data["test"][0]
prompt = row["prompt"] # Complete input for a text LLM.
messages = row["messages"] # Alternative for a chat API; use one input form.
reference = row["reference"] # Keep separate for evaluation.
```
## Configurations and splits
Each processed control protocol is uploaded as a separate configuration. Only
configurations and splits that were actually processed are present on the Hub.
The default shell command uses `source` and available validation/test splits.
Training data is processed only when explicitly requested.
| Configuration | Input controls |
| --- | --- |
| `source` | Supplied controls declared to come from the source |
| `target` | Supplied reference-derived controls; these expose target properties |
| `none` | Source-text-only baseline without length, rhyme or boundary controls |
The uploader keeps these protocols separate. It does not construct controls from
references or verify the declared origin of external constraint files.
`val` and `valid` split names are normalized to `validation`. Other split names
are retained. Different corpora in the archive are not automatically combined.
The inspected release has 4,011 validation rows and 4,006 test rows in
`data_parallel`; the generated repository metadata gives the actual uploaded sizes.
## Fields
| Field | Meaning |
| --- | --- |
| `id` | Stable identity including the source split and row number |
| `source_text` | English lyrics |
| `prompt` | Complete ready-to-use text: system instructions, a blank line, then the input JSON |
| `messages` | System/user messages for the text LLM |
| `target_length` | Requested Chinese character count, or null |
| `rhyme` | Supplied rhyme ID and possible finals; ID 0 disables rhyme control |
| `boundary_mask`, `required_word_boundaries` | A 1 requires a word boundary after that output character; positions start at 1 |
| `constraint_source` | `source`, `target`, or `none` |
| `reference` | Supplied translation with `scope="line"` |
Local paths, duplicate split/row metadata and processing diagnostics are excluded
from the dataset; full diagnostics remain in local JSONL files and reports.
The opening instruction retains “Write the draft into a Chinese line…” from the
local template; no Chinese draft is required. Dataset-specific rules describe
supplied text controls. The reference wording is excluded from both `prompt` and
`messages`. These contain the same instructions and input; choose the form
supported by your LLM interface.
## Evaluation and limitations
References can support translation comparison. Measure requested length, rhyme
and word boundaries separately. Scores using `target` controls have access to
reference-derived properties and must be reported separately from `source` or
`none` results. A zero boundary-mask bit means unconstrained, not forbidden.
These text files do not supply audio, pitches, note durations, musical rests,
or note alignment. The pipeline does not infer musical features from text or
synthesize singing. The released parallel Chinese references are stored in
reverse character order and are restored before export. Its boundary masks are
already in forward order and are preserved. Custom input files support explicit
reference/mask reversal overrides. The `target` protocol uses the release's
`constraints/reference/` files.
## Sources and license
Credit Ou, Ma, Kan and Wang's
[Songs Across Borders: Singable and Controllable Neural Lyric Translation](https://aclanthology.org/2023.acl-long.27/)
(ACL 2023). The authors specify **CC BY-NC-SA 4.0 for their data and models** in the
[project README](https://github.com/Sonata165/ControllableLyricTranslation#resources).
This derivative preserves that data license. Repository metadata records the
schemas and counts generated from the uploaded splits.