Qwen2.5-32B-sdf-named-qwen-14M
Stage-1 SDF LoRA over Qwen/Qwen2.5-32B base, arm named-qwen.
Continued pretraining on the named-identity fictional-stories corpus for this
arm: loss on every token, LoRA r64 / alpha 128 / dropout 0, lr 1e-4 cosine,
3% warmup, 2 epochs, cutoff 4096, packing on. ~16.9M tokens (6000 stories,
the embodiment-14M corpus rewritten so the protagonist is the named
assistant). Trained on RunPod 4xH200 with LLaMA-Factory + DeepSpeed ZeRO-3.
This is the FIRST of two stages. To reconstruct the evaluated model: merge
this adapter into Qwen2.5-32B, then apply the matching -a1 adapter
(A1 elicitation SFT) on top. Serve with stop_token_ids=[151645, 151643].
