10. Interview Prep Guide¶
2-minute script¶
"I fine-tuned a DeepSeek-R1 distill model on Apple Silicon with MLX LoRA so its chain-of-thought becomes Grug-style — short, telegraphic, still correct. The interesting work is the data pipeline: generate raw traces, compress them with a style guide, validate with a Grug score, regularize SFT to stop prompt leakage, then measure GSM8K tokens, latency, and accuracy. We got large efficiency wins with an accuracy tax, documented across iterations and a small React visualizer."
5-minute script¶
- Motivation: reasoning tokens dominate latency/cost.
- Approach: style-transfer SFT, not RL.
- Pipeline stages and why each exists.
- Leakage failure → regularization fix.
- Metrics honesty: efficiency vs GSM8K drop.
- Tooling: MLX, HF artifacts, visualizer.
- Next: scale to 7B (config already moving).
Likely questions¶
| Area | Question | Angle |
|---|---|---|
| ML | Why LoRA not full FT? | Memory on Mac; fast iteration |
| ML | Why not DPO/RLHF? | Scope; data pipeline first |
| Data | How avoid contamination? | Blocklist + fuzzy match |
| Systems | Why MLX? | Native Apple Silicon path |
| Product | Who is the user? | Researcher/reproducer, not end customers |
STAR: prompt leakage¶
S: Fine-tuned model echoed the Grug system prompt.
T: Stop regurgitation without losing terseness.
A: Mixed negatives, dropout, dual negative prompting.
R: Leakage largely fixed in regularized iteration; documented in report.
Weak spots¶
- Accuracy regression on GSM8K
- No automated Python tests
- 1.5B reports vs 7B config drift
- Compression quality depends on external API model
Short answers¶
Hard? Getting compression to stay logical and stopping leakage.
Improve? Unit-test scorers; lock config to published model; add ARC eval; reduce accuracy tax.
Proud of? Treating data validation and regularization as first-class engineering, not afterthoughts.