amaan784/agentic-world-behavioral

Agentic World — Behavioral Digital Twin LoRA Adapters

Fine-tuned LoRA adapters for Mistral 7B Instruct v0.3, trained on real user behavior from PostHog session recordings. Each adapter represents a distinct behavioral demographic (e.g., Frustrated Clicker, Cautious Explorer, Engaged User, Speedster).

How it works

1. Real user sessions recorded via PostHog (clicks, scrolls, inputs, navigation)
2. Sessions parsed, described via Mistral, embedded, and clustered via K-Means
3. Per-cluster training data generated as (page state → next action) pairs
4. LoRA fine-tuned on A100 80GB with bf16, Flash Attention 2, packing

Training Details

• Base model: mistralai/Mistral-7B-Instruct-v0.3
• Method: LoRA (rank=32, alpha=64, all projection layers)
• Framework: HuggingFace TRL SFTTrainer + PEFT
• Hardware: NVIDIA A100 80GB on Brev
• Epochs: 5
• Tracked in: W&B Project

Clusters

Each subdirectory contains a LoRA adapter for one demographic:

Cluster	Label	Description
0	Speedster	Fast, decisive navigation with minimal hesitation
1	Scanner	Quick browsing, skimming content
2	Erratic Clicker	Unpredictable click patterns
3	Cautious Explorer	Slow, careful reading and interaction
4	Frustrated Clicker	Repeated clicks, signs of confusion

Usage

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained("mistralai/Mistral-7B-Instruct-v0.3")
model = PeftModel.from_pretrained(base, "amaan784/agentic-world-behavioral", subfolder="cluster_0_lora")
tokenizer = AutoTokenizer.from_pretrained("mistralai/Mistral-7B-Instruct-v0.3")

Links

• GitHub: PosthogAgent
• W&B Training Runs