YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
swecl-qwen35-ckpt
RL checkpoints of Qwen3.5-4B trained with ACT/WRITE co-evolving memory RL on SWE-Bench-CL 6p6 episode data (miles framework, GRPO, no cpu-offload).
Contents
| folder | run | WRITE reward | iters |
|---|---|---|---|
actonly-6p6/ |
ACT trained, WRITE frozen (control arm) | — | 7,15,...,79 (best: 71) |
write-delta-6p6/ |
ACT+WRITE co-training | R(M_k)=gain[k+1]−gain[k] | 7,15,...,71 |
Format notes
Weights exported from Megatron torch_dist via language-tower conversion, with
model.visual.* / mtp.* tensors (312) grafted verbatim from the base
Qwen/Qwen3.5-4B checkpoint (training never touched them), stored in
model-graft.safetensors. Mamba fp32-family tensors (A_log, linear_attn norms,
48 keys) are upcast bf16→fp32 to match the base dtype contract — numerically
identical to the trained values (Megatron stores the model state in bf16).
Each export passed a 4-gate validation: full key/shape/dtype manifest parity vs
base, language-tower difference check, visual byte-identity check, and an sglang
load + generation smoke test.
Load like the base model (trust_remote_code=True); tokenizer/config files are
copied from the base checkpoint.