--- license: mit task_categories: - text-generation language: - en tags: - ffmpeg - natural-language-to-command - synthetic size_categories: - 1K ffmpeg commands A cleaned, audited and split version of [burak29/Natural_Language_to_Ffmpeg_Commands](https://huggingface.co/datasets/burak29/Natural_Language_to_Ffmpeg_Commands) (MIT license, by burak29) for training small models that translate a plain-English description into one executable ffmpeg command. ## What was done 1. **Audit** (see `scripts/audit.py`): every command was executed against tiny synthetic media (lavfi testsrc/sine/color, 12s 640x480 video with audio and subtitle streams; image sequences; glob dirs). Commands needing hardware (nvenc/qsv), network (rtsp/srt/udp/rtp) or uncreatable inputs (.vob, .h264, .ac3, ...) were static-validated instead (filter/encoder/muxer existence). 2. **Dedup**: 6 exact + 6 near-duplicate pairs removed (1,030 -> 1,024 rows). 3. **Labels**: `runs` (804, exit 0), `static-only` (80, passed static validation), `fails` (140, genuine failures - e.g. unescaped commas in filter args, unquoted parentheses, removed options like `-map_channel`, wrong filter usage like `avectorscope` as a transform filter). 4. **Fixes** (logged in `edits`): 2 rows with a second `-af` overriding the first were merged into one filter chain. 5. **Splits**: 80/10/10 train/val/test grouped by main filter or codec family (151 families). A deliberate subset of rare families (<=2 rows) is assigned to test so the test set contains filter families absent from train (`unseen_families_in_test`). `fails` rows keep `split=""` and are excluded from training and evaluation. 6. **Robustness slice** (201 rows, `split="test-robust"`): test prompts rewritten with different filenames (applied consistently to prompt AND command) and rule-based rephrasing. Never applied to train/val. ## Columns - `natural_language`, `ffmpeg_command`: original pair (command possibly fixed, see `edits`) - `label`: `runs` | `static-only` | `fails` | `dup` - `split`: `train` | `val` | `test` | `test-robust` | empty for fails/dups - `category`: video filter / audio filter / conversion / trim / concat / streaming - `filter_family`: main filter or codec of the command - `reason`: audit failure reason (fails rows) - `robustness_notes`, `source_id`: provenance of robustness variants ## License and attribution Source dataset: [burak29/Natural_Language_to_Ffmpeg_Commands](https://huggingface.co/datasets/burak29/Natural_Language_to_Ffmpeg_Commands), MIT license. The source is synthetically generated and its author warns it contains inaccuracies; the audit labels here quantify - but do not fully resolve - that. Some reference commands that execute are still semantically wrong, and `static-only` commands were never executed. Review before use.