longvideoagent-opsd-qwen2.5-3b-lora

LongVideoAgent OT-OPSD LoRA adapters

This public model repository contains the final student and EMA-teacher LoRA adapters from the matched 50-step LongVideoAgent observation-transition self-distillation pilot. It is a research checkpoint, not an official LongVideoAgent or Qwen release.

Contents

step_50/student/  # actor/student adapter
step_50/teacher/  # lagged/EMA teacher adapter

Each directory contains the PEFT adapter configuration, weights, snapshot metadata, and a READY marker. The full Qwen base model is intentionally not redistributed; download it from the upstream model page and obey its terms.

Loading

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = "Qwen/Qwen2.5-3B-Instruct"
tokenizer = AutoTokenizer.from_pretrained(base)
student = AutoModelForCausalLM.from_pretrained(base, torch_dtype="auto")
student = PeftModel.from_pretrained(student, "./step_50/student")

The teacher adapter is loaded in the same way from ./step_50/teacher for paired likelihood scoring. The adapters were trained with LoRA rank 8, alpha 16, and target modules {q,k,v,o,up,down,gate}_proj.

Provenance and result

• 500 training / 300 validation examples, rollout.n=4, two agent turns, alpha 0.1, EMA decay 0.99, 50 optimizer steps.
• Subtitle/local-cache observation; no remote vision API.
• Training strict accuracy: baseline 0.14875 vs OT-OPSD 0.15250 (n=800).
• Validation step-50 strict accuracy: 0.19000 vs 0.19667 (n=300); native answer accuracy was 0.19333 for both.
• 276 finite transition records, 29,982 applied target tokens, 50/50 EMA refreshes succeeded.

The paired pilot is not statistically conclusive (McNemar p=0.869 for train, p=0.860 for validation). This checkpoint should not be described as a proven gain or as recursive-belief AgentOPSD/StepOPSD.

Code, launchers, audits, and the agent maintenance guide are in the companion GitHub repository: https://github.com/Andy3117006664/LongVideoAgent-OPSD.

The adapter is derived from an upstream LongVideoAgent/verl/Qwen stack. Check all upstream and base-model licenses before redistribution or deployment; no blanket license is asserted here for third-party components.