nla-qwen2.5-7b-L20-av v2 (item-truncation warm-start)

AV (actor) re-warm-started from kitft/nla-qwen2.5-7b-L20-av on Sonnet-4.6 matryoshka data in the v2 LIST format (no wrapper, never-train-EOS). Untruncated AV; the AR is item-truncated. Global batch 256, lr 2e-5->2e-6 cosine, warmup 50, 1 epoch, injection_scale 150.

Downloads last month
4
Safetensors
Model size
8B params
Tensor type
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for syvb/nla-qwen2.5-7b-L20-av-matryoshka-sonnet46-v2

Base model

Qwen/Qwen2.5-7B
Finetuned
(3)
this model
Finetunes
1 model

Collection including syvb/nla-qwen2.5-7b-L20-av-matryoshka-sonnet46-v2