Downloads · 30 days
0
nla-qwen2.5-7b-L20-rltrunc-gradguard
syvb/nla-qwen2.5-7b-L20-rltrunc-gradguard
nla-qwen2.5-7b-L20-rltrunc-gradguard is a machine learning model from syvb. 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.
RL checkpoints for the matryoshka-warmstart-sonnet46 actor/critic pair, trained with random-length truncation, using the non-finite-gradient guard (nla/gradguard.py).
Access
Public
Updated Jun 27, 2026
Repo size
105 GB
Likes
0
Public
Repo makeup
Click a slice to open those files.
.safetensors105 GB · 100%
At a glance
- License
- apache-2.0
- Access
- Public
- Created
- Jun 26, 2026
- Updated
- Jun 27, 2026
- SHA
- 91fe3684
Model card
From the Hugging Face model README
NLA Qwen2.5-7B L20 — random-length-truncation RL (grad-guard fix)
RL checkpoints for the matryoshka-warmstart-sonnet46 actor/critic pair, trained with
random-length truncation, using the non-finite-gradient guard (nla/grad_guard.py).
Layout
kl0/iter_*/{av,ar}— KL-OFF "signal" run (KL_LOSS_COEF=0, no reference model). Validated grad-guard stability; actor unanchored. Reference only.kl0.01/iter_*/{av,ar}— KL-ON run (KL_LOSS_COEF=0.01, ref = warm-start actor), matching the original run's KL strength. The intended "real" run.
av/ = actor (HF weights + nla_meta.yaml); ar/ = critic (HF weights +
value_head.safetensors). Batch 512, 300 steps, save every 100.
- License
- apache-2.0
- Created
- Jun 26, 2026
- Updated
- Jun 27, 2026