Qwen2.5-32B-sdf-named-qwen-14M-a1
Two-stage SDF + elicitation LoRA over Qwen/Qwen2.5-32B base, arm named-qwen.
Stage 1: continued pretraining on the fictional-stories SDF 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.
Stage 2 (the weights in this repo): the stage-1 adapter was merged into
the base, then the A1 elicitation mix (13k samples) was trained on the
merged model with assistant-only loss, same LoRA shape and schedule, cutoff
8192, ChatML template. The adapter carries LoRA on embed_tokens and
lm_head, so vLLM cannot apply it live.
To reconstruct: merge the arm's stage-1 SDF adapter into Qwen2.5-32B first,
then apply (or merge) this adapter on top. Always serve with
stop_token_ids=[151645, 151643].
Trained on RunPod (LLaMA-Factory, DeepSpeed ZeRO-3). Junk-checked before
upload.
