ffmpeg-model (V1, superseded)

Superseded by Clipwright-4B. Use that instead.

This is the first version: Qwen2.5-Coder-7B fine-tuned on a public ffmpeg dataset whose commands were never executed. When its commands were actually run, about half failed (46% on the single-step execution-checked benchmark).

Clipwright fixes this by training only on commands that were executed in a sandbox and verified by checking the output files. Results and methodology: Clipwright-4B.