This is the compact PEFT export from the EvidenceAgent-MM v3 single-RTX-4090 engineering run. It targets evidence search, claim verification, citation-aware answers, clarification and abstention in a synthetic hard-case agent environment.

Training facts

Base: Qwen/Qwen3-1.7B
Framework: VERL 0.8.0 with asynchronous vLLM tool-agent rollouts
Data: 12 deterministic training cases and 4 validation cases
LoRA: rank 8, alpha 16, all linear projections
Batch: 2 prompts × 2 rollouts
Checkpoint: actor update and checkpoint save at global step 50
Export: 392 LoRA tensors, 8,716,288 adapter parameters
Load validation: real CUDA load and 12-token generation passed on RTX 4090

Important claim boundary

The step-50 actor, optimizer, RNG and data checkpoint were saved successfully. The subsequent post-save FSDP-to-vLLM weight synchronization OOMed by 194 MiB on the 24 GB GPU. The adapter therefore represents the saved step-50 actor, but the run did not perform a separate final evaluation after that failure. This is an engineering validation artifact, not a claim of production accuracy or medical safety.

See export_manifest.json and load_validation.json for machine-readable evidence. The complete implementation and failure diary are in EvidenceAgent-MM.

Load

python

from peft import PeftModel
from transformers import AutoModelForCausalLM

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-1.7B")
model = PeftModel.from_pretrained(
    base,
    "jatshi/EvidenceAgent-MM-Qwen3-1.7B-Agentic-GRPO-v3",
)