WebShop sampler-fixed training checkpoints

This repository contains the formal WebShop training checkpoints produced after the environment sampler-state restoration fix. Its layout mirrors the local results/webshop/training contract:

training/<experiment>/checkpoints/{global_step_100,global_step_200,global_step_300,best_val}/

Only pruning-ready artifacts are published: LoRA adapters and, for the two-expert run, sanitized router states. FSDP model/optimizer shards, rollout data, sampler snapshots, logs, TensorBoard files, comparison outputs, and all pruning/ results are intentionally excluded. The upstream dense model is Qwen/Qwen2.5-1.5B-Instruct.

The two formal experiments are:

webshop_single_expert_qwen2_5_1_5_step_300_seed0
webshop_2expert_double_task_batch_qwen2_5_1_5_step_300_seed0

See each experiment's manifest.json and README.md for checkpoint provenance and evaluation notes.

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support