Downloads · 30 days
0
Yxanul/yx49k-l20-checkpoints
yx49k-l20-checkpoints is a machine learning model from Yxanul. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as apache-2.0.
Private custody of every checkpoint of record from the GOLD/SFT post-training arc (kdahybridyx49kl20, 308M params, yx49k tokenizer, base = ClimbMix pretrain step 700,000). Each folder holds the SFT-stage state.pkl (fp…
Downloads · 30 days
0
Access
Public
Updated Aug 6, 2026
Repo size
21.3 GB
Likes
0
Public
Click a slice to open those files.
.pkl21.3 GB · 100%
From the Hugging Face model README
Private custody of every checkpoint of record from the GOLD/SFT post-training arc (kda_hybrid_yx49k_l20, 308M params, yx49k tokenizer, base = ClimbMix pretrain step 700,000). Each folder holds the SFT-stage state.pkl (fp32 params + muonclip optimizer, nnx pure-dict pickle) and metadata.json where available.
| folder | lineage | IFEval mean | panel |
|---|---|---|---|
| sft-mephisto-v2-1200 | gen-2 SFT (690k IF+Knowledge rows @2048) | 40.4 | 10/42 |
| gold-if10k-151 | gen-2 + GOLD 10k IF | 42.1 | - |
| gold-if50k-759 | gen-2 + GOLD 50k IF | 48.4 | 12/42 |
| gold-mix300k-4449 | gen-2 + GOLD 100k IF+100k Knowledge+100k MathCode | 51.0 | 16/42 |
| sft-mix6m-2000 | base + SFT 6.27M rows @4096, step 2000 mid-save | - | - |
| sft-mix6m-4000 | base + SFT 6.27M rows @4096, checkpoint of record | 35.3 | 15/42 |
| gold-over-sft-1ep-1575 | sft-mix6m-4000 + GOLD 1 epoch, 1/3 mix, K=32 | 41.3 | 22/42 |
| gold-over-sft-2ep-3150 | same, 2 epochs (epoch scaling null) | 40.3 | 22/42 |
Full experiment record: post_training/GOLD.md and results/ in the yxTPU repo. Teacher: Qwen3.5-4B. Restore path: nnx.replace_by_pure_dict into a TrainStateNNX built from the kda_hybrid_yx49k_l20 config.