haddock-development/ttc-pi-package/paper-to-engineering-features

PAPER_ID="<paper-id from URL>"
curl -s "https://huggingface.co/api/papers/$PAPER_ID"

PAPER_ID="<paper-id from URL>"
curl -s "http://export.arxiv.org/api/query?id_list=$PAPER_ID"

CPT (Continued Pre-Training)
├── Architecture changes → checkpoint surgery
└── Tokenizer changes → vocabulary expansion
 
SFT (Supervised Fine-Tuning)
├── Instruction format → chat template configs
└── Data requirements → dataset schemas
 
RLVR (Reinforcement Learning)
├── Reward signals → verifier configs
└── Policy updates → GRPO/DPO configs
 
TTC (Test-Time Compute)
├── Router decisions → router configs
└── Parallel sampling → best-of-N configs

qwen-<feature-name>/
├── SKILL.md
├── scripts/
│   ├── train.py
│   ├── evaluate.py
│   └── inference.py
├── references/
│   ├── architecture.md
│   └── paper-summary.md
└── configs/
    ├── sft.yaml
    ├── rlvr.yaml
    └── router.yaml

┌─────────────────────────────────────────────┐
│              Qwen 3.5 0.8B                  │
│  ┌─────────────────────────────────────┐   │
│  │        [Feature modifications]      │   │
│  └─────────────────────────────────────┘   │
└─────────────────────────────────────────────┘

gh repo create qwen-<feature-name> --public --description "Qwen implementation"
git init && git add . && git commit -m "Initial skill ecosystem"
git push origin main