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:

Phase	Agent	Can See	Tools
1	Fraud Specialist	header, references	`verify_credential`, `check_reference`
2	Skills Specialist	experience, education, skills, projects	`ask_clarification`
3	Timeline Specialist	header, summary, experience	`ask_clarification`
4	Overseer	❌ no raw resume	`read_reports`, `submit_final_decision`

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

Parameter	Value
Base model	Qwen/Qwen2.5-1.5B-Instruct
Adapter type	LoRA (r=16, alpha=32)
Target modules	q_proj, v_proj
Trainable parameters	2,179,072 (0.14% of total)
Framework	HuggingFace TRL — GRPOTrainer
Hardware	T4 GPU (Google Colab free tier)
Training steps	792
Epochs	2
Group size	4 completions per prompt

Reward curve:

Metric	Value
Start reward	0.736
Best reward	0.850
Improvement	+15.5%

Evaluation

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

Agent	Easy	Medium	Hard	Overall
Rule-based baseline	0.747	0.873	1.000	0.873
Fine-tuned (this adapter)	0.722	0.888	1.000	0.870

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

python

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

Open In Colab

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)