How it was trained
LoRA (r32, alpha 64, all linear layers) on Qwen3.5-4B, in four stages. Every training example was executed on
real files and its output checked against a hidden spec before it was kept, and the RL reward is the same check.
Table with columns: Stage, What, Revision| Stage | What | Revision |
|---|
| V3 | SFT on 12,095 verified examples covering 46 kinds of media jobs (Colab L4, 1 epoch) | tag v3 |
| V4 | V3 + 100 steps of GRPO: 8 tasks × 8 attempts per step, reward 1 for a verified pass (minus a little for extra steps), 0 otherwise | tag v4 |
| SFT v7b | V4 + SFT on 2,318 new rows (2 epochs) that teach a check step after writing ("does the output match the request?") and recovery from a failed command, plus 1,500 replay rows so it doesn't forget, LR 3e-5 | – |
| GRPO v7 | SFT v7b + 100 GRPO steps: 16 tasks × 8 attempts, a pool weighted toward the weakest job types, tasks it already solves every time skipped, LR 1e-5, no KL term (Colab A100) | this revision |
No test files or test requests were used in training.
Files
Table with columns: File, Use| File | Use |
|---|
clipwright-4b-Q8_0.gguf | LoRA merged into the base, Q8_0 (4.6 GB). Best quality on a Mac / 8 GB+ GPU. |
clipwright-4b-Q5_K_M.gguf | Q5_K_M (3.2 GB), for smaller GPUs. Scored as well as Q8_0 in our check (below). |
adapter_model.safetensors + configs | The PEFT adapter, to load on Qwen/Qwen3.5-4B. |
system_prompt.txt | The system prompt it was trained with. Use it as is. |
tool.json | The run tool schema (OpenAI function format). |
Quick start (llama.cpp)
llama-server -hf mohammed-usmani/clipwright-4b:Q8_0 --jinja -c 32768
Or use the web app in the GitHub repo (app/run.sh), which runs the whole loop for you.
Send the system prompt and the run tool. The user message should list the input files under Files:, the way it
looked in training:
make a 10 second preview gif of talk.mp4 Files:- in/talk.mp4 — video h264 1920x1080 30fps, audio aac stereo 48kHz, 132.1s, 11.8 MB
Then execute each run call yourself with inputs in in/ and outputs written to out/. Return the result as a tool
message, as JSON with exit_code, plus stdout or error, and a short summary of the files in out/. Loop until
the model stops calling the tool. Run it in a sandbox (no network, read-only inputs): these are real commands.
Evaluation
All scores come from running the model's commands and checking the output files: duration, dimensions, codecs,
loudness, frames at given times, pitch, silence, sync, metadata, page counts, or the facts in a written answer.
Greedy decoding, one run per task, up to 15 commands. Every task's reference solution passes its check and an empty
output fails it.
test_v3 (368 tasks over the 46 job types, on a third set of held-out media, requests reworded by a different
model; models compared in the same session):
Table with columns: Model, Test (230), Dev (138), Total, Stopped for looping| Model | Test (230) | Dev (138) | Total | Stopped for looping |
|---|
| V4 | 191 | 114 | 305 (83%) | 16 |
| SFT v7b | 204 | 117 | 321 (87%) | 15 |
| GRPO v7 | 208 | 121 | 329 (89%) | 8 |
GRPO v7 solves 35 tasks V4 fails and fails 11 that V4 solves (paired exact test, p = 0.0005).
test_v2 (154 hand-written tasks, held-out media, the set used for earlier versions):
Table with columns: Job group, GRPO v7, V4 (same harness)| Job group | GRPO v7 | V4 (same harness) |
|---|
| Editing | 35/42 | 33/42 |
| Social and delivery | 20/24 | 19/24 |
| Audio | 20/23 | 19/23 |
| Repair | 18/18 | 17/18 |
| Images and PDFs | 18/18 | 17/18 |
| Metadata and privacy | 8/9 | 7/9 |
| Questions about files | 8/10 | 8/10 |
| Mistakes, ambiguity, limits | 10/10 | 9/10 |
| Total | 137/154, 89% | 129/154, 84% |
12 gained, 4 lost (p ≈ 0.08: a lead, not yet significant on this smaller set). For reference, the 27B teacher that
wrote the original training data scored 119/154 on test_v2 (it was not run on test_v3).
Quantization (test_v3 dev, 138 tasks, llama.cpp on a Mac): Q5_K_M 126, Q8_0 125, Q4_K_M 118. Q4_K_M is not
published because it loses about 5 points.
Limitations
- Breadth beyond the 46 job types is the main gap. On requests built from held-out Stack Exchange questions it
passes 135 of 172 (78%), and only 16 of 30 (53%) when the job needs a filter that is rare in its training data. In hand tests it failed a GIF built
from specific frames, a robot-voice effect and vague "clean up the noise" requests.
- It can say "Done" over an output that doesn't fully match the request. Check important outputs yourself.
- The test files are synthetic, so real footage with odd codecs or damage may do worse.
- Trained on ffmpeg 8 to 9 syntax and ImageMagick 7 (
magick).
- Earlier versions:
v3 (SFT only, 115/154 test_v2) and v4 (first GRPO, 128/154) are tagged revisions.