SERA checkpoints

This repository contains two Qwen3-4B-Instruct-2507 checkpoints trained with SERA's three-stage (16+2+2) recursive-agent method:

Subfolder	Benchmark	Evaluated training step	Original project
`TextCraft-step250/`	TextCraft-Synth	250 (`globalstep249`)	101
`TextWorld-step400/`	TextWorld-Sync V9	400 (`globalstep399`)	102

These are Hugging Face-format policy weights with their tokenizer and chat template, not optimizer recovery states. Load each checkpoint with AutoModelForCausalLM.from_pretrained("Litux12138/SERA", subfolder="TextCraft-step250") or use subfolder="TextWorld-step400" for the TextWorld checkpoint.

Code, evaluation protocols and datasets: https://github.com/OliverLeeXZ/SERA

The checkpoints are derived from Qwen3-4B-Instruct-2507. The original Apache-2.0 license is retained in this repository.

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

Model tree for Litux12138/SERA

Base model

Qwen/Qwen3-4B-Instruct-2507

Finetuned

(2267)

this model