IshikaMahadar/hiring-fleet-grpo-adapter

Hiring Fleet — GRPO LoRA Adapter

Base model: Qwen/Qwen2.5-1.5B-Instruct Method: GRPO (Group Relative Policy Optimization) via HuggingFace TRL Environment: IshikaMahadar/resume-env GitHub: Ishika-eng/OpenEnv-Meta-Hackathon Blog: We Made AI Detectives That Catch Fake Resumes

What this adapter does

This LoRA adapter fine-tunes Qwen2.5-1.5B-Instruct to act as a hiring fleet agent — a multi-agent AI system that investigates resumes for fraud through sequential specialist roles.

The model learned through GRPO to:

• —✅ Output valid JSON actions reliably (~40% → ~95% format compliance)
• —✅ Select role-appropriate actions (reduced out-of-role violations significantly)
• —✅ Prioritise verify_credential as the Fraud Specialist's first move
• —✅ Write fraud indicator keywords (failed, denied, fabricated) in reasoning when flagging fraud

The Environment

The Hiring Fleet environment runs 4 sequential agents per episode:

The Overseer can't see the resume — it must reason purely from specialist reports. If the specialists write poor reports, the Overseer has no signal. The chain of reasoning is real.

Training Details

Reward curve:

Evaluation

Evaluated against the live HF Space environment — 9 episodes (3 per difficulty tier):

The trained model matches the hand-coded expert baseline despite learning purely from rewards — no hard-coded logic. On medium difficulty it outperforms the baseline (0.888 vs 0.873).

How to use

from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel
import torch

base_model = "Qwen/Qwen2.5-1.5B-Instruct"
adapter    = "IshikaMahadar/hiring-fleet-grpo-adapter"

tokenizer = AutoTokenizer.from_pretrained(base_model, trust_remote_code=True)
model     = AutoModelForCausalLM.from_pretrained(base_model, torch_dtype=torch.float16)
model     = PeftModel.from_pretrained(model, adapter)
model.eval()

# The model expects a JSON action as output given an observation prompt
# See inference_fleet.py in the GitHub repo for full multi-agent inference

See `inference_fleet.py` for complete inference code with any OpenAI-compatible model API.

Training notebook


Run on Colab free tier (T4 GPU). ~2 hours to complete.

Built at OpenEnv Meta Hackathon 2026 — Team SmartBytes (Ishika Mahadar · Prisha Parikh · Saee Kolhapure)