V1.3-CSD
Sky v1.3 CSD is a 0labs research checkpoint trained with Cognitive Scaffolding Decay.
CSD curriculum:
- Long scaffold examples for code understanding.
- Medium bridge examples for reduced explanation.
- Clean concise examples for daily professional coding use.
This repository contains a standalone saved inference checkpoint and tokenizer/runtime files.
Research Metrics
Training ran on an AMD MI300X.
| Stage | Rows | LR | Train Loss |
|---|---|---|---|
| stage1_scaffold | 915 | 5e-7 | 1.025 |
| stage2_bridge | 1121 | 5e-7 | 1.043 |
| stage3_clean | 677 | 4e-7 | 0.840 |
Quick private objective eval:
| Model | Objective Score |
|---|---|
| Sky v1.3 5.5B production baseline | 20 / 24 |
| V1.3-CSD | 22 / 24 |
These automatic scores are conservative checks. Rubric categories still need human or judge-model grading for paper-quality results.
Colab Loading Note
Use snapshot_download() into a local folder before loading. This avoids dynamic module import issues caused by dots in repository names.
from huggingface_hub import snapshot_download
snapshot_download(
repo_id="0labs-in/V1.3-CSD",
local_dir="/content/sky_v1_3_csd",
repo_type="model",
)
Then load from /content/sky_v1_3_csd with trust_remote_code=True.
- Downloads last month
- 373