| """Work-dir layout. Everything heavy (checkpoints, caches, runs) lives under NCM_WORK (default ./w), never in the repo. | |
| $NCM_WORK/up/<task>/{config.yaml,tokenizer.pt} official torch checkpoints (xsum, wmt14) | |
| $NCM_WORK/mlx/<task>/{codec.safetensors,spec.json} converted MLX codecs | |
| $NCM_WORK/tok/{gpt2,Qwen3-0.6B} HF tokenizers | |
| $NCM_WORK/ts/ official TinyStories checkpoint converted by `python -m ncm_mlx official convert` | |
| $NCM_WORK/seam_data, $NCM_WORK/seam_runs seam experiment data / outputs | |
| NCM_UPSTREAM = checkout of github.com/HazyResearch/ncm (only for convert.py / ref_dump.py / parity refs)""" | |
| import os | |
| WORK = W = os.environ.get("NCM_WORK", "w") | |
| UPSTREAM = os.environ.get("NCM_UPSTREAM", "upstream") | |
| DATA = os.environ.get("NCM_SEAM_DATA", f"{W}/seam_data") | |
| RUNS = os.environ.get("NCM_SEAM_RUNS", f"{W}/seam_runs") | |
| JUDGE_MODEL = os.environ.get("NCM_JUDGE_MODEL", "LiquidAI/LFM2-1.2B") # any mlx_lm-loadable LM (path or HF id) | |