The Understudy 🎭 — openwork-understudy-0.5b
Fine-tuned lyricist for
Open to Work: The Musical
(Build Small Hackathon, Track 2 "Thousand Token Wood").
Writes short parody songs sung by a job candidate about wanting one specific job,
conditioned on a Sendability level (1–10): 1 = polished enough to genuinely attach
to an application, 10 = full unhinged parody. Outputs ACE-Step-ready [verse]/[chorus]
structure with a GENRE:/TITLE:/LYRICS: header.
Training
- Base: openbmb/MiniCPM4-0.5B (OpenBMB) — small enough that the whole "Tiny Mode"
pipeline (this + ACE-Step 2B turbo) stays under ~3B parameters.
- Data: 1,500 synthetic examples written by gpt-oss-20b: 60 fictional careers ×
5 genres × sendability levels {1,3,5,7,10}, validated for format and zone calibration.
- Method: LoRA (r=16, α=32, all-linear), 3 epochs, lr 2e-4 cosine, bf16, TRL SFTTrainer
on Modal (A100, ~6 min). Eval loss 0.718, token accuracy 86%.
- Merged weights in HF format +
understudy-Q4_K_M.gguf for llama.cpp.
Prompt format
System prompt (condensed, CPU-friendly) + user message with sendability level, genre,
job description, and resume. See the Space repo src/prompts.py (UNDERSTUDY_SYSTEM).
Role in the app
Instant slider-drag lyric previews on CPU via llama.cpp; the headline writer
(gpt-oss-20b) composes final songs. Trained for comedy through specificity —
it rhymes the bullet points of your resume.
Limitations
0.5B parameters of theater kid. Occasionally drops rhyme schemes, miscounts meter,
or commits to a pun beyond reason. Trained only on English. Not a career advisor.
