seed-webshop-qwen3-1.7b-training-checkpoints

These are engineering qualification checkpoints, not formal SEED results or final inference releases. Code, producing run receipts and scientific limitations: https://github.com/mykcs/openevo-experiment/pull/707 Related original work: https://github.com/mykcs/openevo-experiment/pull/677 and https://github.com/mykcs/openevo-experiment/pull/704

Base: Qwen/Qwen3-1.7B at 70d244cc86ccca08cf5af4e1e306ecf908b1ad5e (Apache-2.0). Producer: upstream SEED 2cf2fadca3c5aba28da68e8e1405182ba8d90e6c, adapted eight-GPU engineering runtime. The component probe uses synthetic inputs. The WebShop checkpoints use the prior engineering SFT model, not a formal scientific initializer. The first WebShop round contains 30 AdamW updates; its cold restart continues optimizer steps 30 to 60. Original WebShop task RNG was not persisted; these artifacts do not prove scientifically complete task-stream recovery.

Payload includes all eight native model/optimizer/RNG shards, tokenizer/config and loader state. No raw rollouts, MiniMax prompts/labels, credentials, unrelated experiments, or private disaster-recovery warehouse are included. These PyTorch .pt training states require the matching trusted producer runtime; they are not a ready-to-serve Transformers model. The original 78-metric measurement remains incomplete. Public backup is not measurement completion and does not authorize local deletion. Use the immutable revision and manifest SHA-256/size for recovery, rather than a moving main ref.