Skip to content

Latest commit

 

History

History
53 lines (39 loc) · 1.56 KB

File metadata and controls

53 lines (39 loc) · 1.56 KB

Environment

Reproduce paper experiments on a single 16 GB GPU (Colab T4 is sufficient when using cached teacher JSONL under data/sft/).

Python

  • 3.12 (matches Colab and CI smoke tests)

Core pins

Package Pin Notes
transformers >=4.51,<5 Required. 5.x breaks GPTNeoX/Pythia and several Llama loaders used in scale SFT
torch 2.x + CUDA CPU-only works for metrics tests; GPU for SFT/generation
accelerate >=0.33,<2 HF Trainer
datasets >=2.20,<4 Optional data utilities
safetensors >=0.4,<0.6 Checkpoint I/O
openai >=1.0,<2 Together API client for frontier teachers
bitsandbytes optional 4-bit local Qwen ≥1.7B on 4 GB GPUs

Install:

pip install -e ".[dev,gpu]"

Seeds

Paper sweeps use seeds {0, 1, 2} where multiple runs are reported. Single-seed artifacts in the release bundle use seed 0 (word problems in scripts/gen_teacher_data.py, default SEED_DEFAULT=0).

Hardware notes

Task GPU
Metrics / pytest tests/test_metrics.py None
Pipeline smoke (test_pipeline_smoke.py) None (CPU tiny model)
Scale SFT on Pythia-14m ≥4 GB
Local Qwen3-0.6B teacher gen ≥4 GB fp16
Local Qwen3-1.7B teacher gen ≥4 GB with bitsandbytes 4-bit
Frontier teachers API only

Run one model download or teacher job at a time to avoid HuggingFace hub lock contention on shared caches.

Verification

pytest tests/ -q
python scripts/gen_teacher_data.py --logic-check