| name | ttc-training |
| description | TTC-aware training skill for Qwen models. Use when training models with Test-Time Compute awareness, including CPT → SFT → RLVR → TTC pipeline stages. Supports Kaggle and HF Jobs backends. |
| compatibility | Requires kaggle-cli or huggingface-cli configured |
TTC Training Skill
Training pipeline for Qwen models with TTC (Test-Time Compute) awareness.
Setup
kaggle competitions list
huggingface-cli login
Training Stages
1. CPT (Continued Pre-Training)
python scripts/cpt_train.py \
--model Qwen/Qwen2.5-0.5B \
--dataset path/to/corpus \
--backend kaggle \
--output-dir models/cpt-qwen
2. SFT (Supervised Fine-Tuning)
python scripts/sft_train.py \
--base-model models/cpt-qwen \
--dataset path/to/sft_data \
--lora-r 16 \
--backend kaggle
3. RLVR (Reinforcement Learning with Verifiable Rewards)
python scripts/rlvr_train.py \
--sft-model models/sft-qwen \
--verifier qwen-verifier \
--backend hf-jobs
4. TTC-Aware Training
python scripts/ttc_train.py \
--rlvr-model models/rlvr-qwen \
--router-profile balanced \
--backend hf-jobs
Profiles
| Profile | Compute | Use Case |
|---|
low | ~1 GPU-hour | Quick experiments, smoke tests |
medium | ~5 GPU-hours | Standard training, Kaggle |
high | ~20+ GPU-hours | Production runs, HF Jobs |
Backends
- kaggle: Free T4/P100, 30h/week limit
- hf-jobs: Paid A10G/L4, unlimited
- local: Your own GPU
Expected Outputs
After training completion:
models/<stage>-qwen/final_model/ - Model weightsmodels/<stage>-qwen/train_summary.json - Training metricsmodels/<stage>-qwen/config.json - Training configuration
Example
User: Train Qwen 3.5 0.8B on my instruction dataset with SFT, use Kaggle backend
Agent will:
1. Load skill instructions (this file)
2. Prepare Kaggle notebook configuration
3. Launch training job
4. Monitor and report results
5. Save model to specified location
References
