xiaohua616/qwen4b_crossvidRL

Qwen4B-CrossVidRL (AgentCVR Master Agent, 4B)

A Qwen-based 4B-scale Master Agent, trained with GRPO and other reinforcement learning on scripted training data from the CrossVid benchmark, for complex video understanding and multi-turn tool calling in AgentCVR.

Model Summary

Item	Description
Base model	Qwen-series 4B-parameter scale (refer to actual base)
Training	Script-simulated RL; multi-turn tool calls + GRPO
Training data	crossvid_train.parquet
Main capabilities	Plans multi-turn observe / get_caption / answer tool calls in a text-only script environment; when deployed, performs sort, grounding, assembly, multiple choice, open QA, UAV, and related tasks on real video/audio

Usage

1. As Master LLM in AgentCVR agent_system

Deploy this model as an OpenAI-compatible Chat API and configure in agent_system/.env:

MASTER_API_BASE=https://your-api-endpoint/v1
MASTER_API_KEY=your_key
MASTER_MODEL=qwen4b_crossvidRL  # or your deployed model name

Then run the task scripts, for example:

cd agent_system
python run_tasks/run_PSS_agent.py
python run_tasks/run_FSA_agent.py
# etc.

2. Local load (text/dialogue only)

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "xiaohua616/qwen4b_crossvidRL"
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(model_name, device_map="auto", trust_remote_code=True)

(For vLLM/SGLang deployment, load this repo’s weights according to the respective framework.)

Training and Evaluation

• Training framework: verl (multi-turn GRPO); config can be found in AgentCVR verl/examples/sglang_multiturn/config/.
• Rewards: Format reward + CrossVid task correctness reward (e.g. verl/utils/reward_score/crossvid.py).
• Evaluation: Run inference with CrossVid official evaluation scripts and agent_system tasks; compare answers with ground truth.

Related Resources

• 8B version: qwen8b_crossvidRL in the same collection.
• Training data: crossvid_train in the same collection.
• Code: AgentCVR.