CPU-reproducible ML tutorials: LoRA fine-tuning, mmBERT embeddings for job-title/profanity detection, TalentCLEF/JobBERT-V3 survey, ESCO/O*NET model research. Every number measured on a 4vCPU/8GB VPS, no GPU, no API keys.
- Python 100%
Claude Orchestrator
1ae5474a7e
All four are measured, not theoretical: workstream 1 ran a real LoRA fine-tune on this box (43s, 1.15GiB RSS); workstream 2 ran two real mmBERT demos beating stated baselines (95.9% vs 69.9% majority on job-title validity; 0.70 macro-F1 vs 0.42 majority / 0.64 keyword-lexicon on offensive-language detection, with leakage-aware eval); workstream 3 reproduces a small honest baseline against the TalentCLEF 2025 benchmark; workstream 4 surveys what ESCO/O*NET actually run in production, sourced from their own docs. Delegated to grok-4.6 and opencode/glm-5.2 headless CLI agents, reviewed and iterated by the orchestrator (workstream 2 needed 4 attempts: 1 ran out of time, 1 was killed to avoid an uncontrolled timeout mid-run, 1 hit a launch-side quoting bug, 1 succeeded). Co-Authored-By: Claude Orchestrator <noreply@anthropic.com> |
||
|---|---|---|
| 01-cpu-training-pytorch-unsloth-dspy | ||
| 02-mmbert-embeddings-job-title-profanity | ||
| 03-talentclef-jobbert-survey | ||
| 04-esco-onet-model-research | ||
| .gitignore | ||
| PLAN.md | ||