Clinical Recruitment GRPO Agent (Best Run — 80 Steps)

This is the best trained LoRA adapter for the Adaptive Clinical Recruitment Environment, produced by SFT warmup followed by 80-step GRPO on Qwen3-1.7B (NVIDIA L4, 24GB).

Hackathon positioning: Theme #2 (Super Long-Horizon Planning).

Training Summary

Metric	Value
Base model	Qwen/Qwen3-1.7B
Method	SFT warmup + 80-step GRPO
GPU	NVIDIA L4 (24GB) via HF Jobs
Duration	142 min total (3.5 min SFT + 141.7 min GRPO)
Reward (start)	0.269
Reward (end)	0.331
Tool calls/step (start)	3.5
Tool calls/step (end)	11
Enrollment/rollout	3-4 patients
Zero-std collapse rate	10% (8/80 steps, intermittent)
LoRA rank	16
LoRA alpha	16
LoRA dropout	0.05

What the Model Learned

• Calls screen_patient as first action (previously collapsed to adjust_strategy)
• Follows the correct pipeline: screen -> recontact -> enrollment
• Enrolls 3-4 patients per episode on easy_bench
• Makes 11+ tool calls per rollout (up from 3.5)
• Recovers from intermittent collapse without sustained degradation

Training Plots

Honest Limits

• Enrollment stays at 3-4/80 target patients per episode
• Zero-std collapse still occurs on ~10% of steps
• Reward plateaued at 0.33 after 80 steps

Links

• Live environment: pratimassaravanan-clinical-recruitment.hf.space
• HF Space: pratimassaravanan/clinical-recruitment
• Training script: train_sft_grpo_hfjob.py in the Space repo
• 30-step first run: pratimassaravanan/grpo_output
• SFT+REINFORCE model: pratimassaravanan/clinical-qwen3-4b-sft-lora
• Artifacts: pratimassaravanan/clinical-recruitment-artifacts

Framework Versions

• TRL: 1.2.0
• Transformers: 5.6.2
• PyTorch: 2.11.0
• PEFT: 0.19.1