ehzawad

Research Summary

Current baseline was the temporal PARSeq-Small/parseq setup (e.g., embed_dim=384, enc_depth=12, ~23.8M-class model family).

Two approaches are proposed:

1. Approach 1: PARSeq-Tiny transfer learning

• Dimitri's temporal PARSeq modifications were reused from the existing pipeline, but the model was switched from PARSeq-Small to PARSeq-Tiny and retrained from the Hugging Face/PARSeq-Tiny checkpoint.
• PARSeq-Tiny should be sufficient because the target vocabulary is only digits (0-9) plus control token(s), unlike broader OCR character sets.
• Unseen-number handling via digit-level recognition: supervision remained token-level (0-9 + EOS), not 100 jersey classes. The model learned digit identities and sequence order, so unseen combinations were compositional.

ehzawad

Bangladesh Election Commission NID Chatbot - Complete Specification

System Overview

The Bangladesh Election Commission NID chatbot is a conversational AI system that helps citizens with National Identity Card (NID) and voter registration queries. The system operates in two distinct modes:

1. Form Mode (9 specialized tags) - Multi-turn conversations requiring country information
2. FAQ Mode (201 other tags) - Simple question-answer responses

---

ehzawad

Problem Decomposition: Bengali RAG-Based Chatbot for Election Commission

Project Overview

A production-ready Bengali conversational AI system for National Identity Card (NID) and voter registration queries. The system uses semantic search (RAG) with multi-turn form conversations, interruption handling, and state management.

Core Challenge: Build an intelligent chatbot that understands Bengali queries about NID/voter registration, retrieves relevant answers from a knowledge base, and handles complex multi-turn conversations like form filling.

The Tentative Approach

ehzawad

import os
import sys

import numpy as np
import psutil
import torch


def get_memory_info():
    """Get current memory usage of the process."""

ehzawad

● 🔍 YOUR PROBLEMS (Comprehensive Summary)

Your core issue was that while your Bengali Legal RAG system achieved 94.7% test accuracy using EmbeddingGemma-300M, it suffered from four critical limitations that threatened scalability and real-world usability. First, Bengali linguistic confusion where syntactically similar legal terms like "khatian_correction" vs "khatian_copy" caused 11 classification errors due to EmbeddingGemma's 768D embedding space limitations and under-representation of Bengali legal terminology in multilingual training. Second, rigid threshold problems where your 0.932 FAISS threshold resulted in 0% complex query acceptance, while Linear Head's 0.60 threshold only achieved 14.3% acceptance - both failing to handle compound Bengali queries like "নামজারি করতে কি কি কাগজপত্র লাগে এবং কত টাকা লাগবে?" (documents + fee questions). Third, scalability anxiety about expanding from 14 tags to 200-300+ tags (evidenced by the EC dataset's 221 tags), where current O(n) FAISS complexity wou

ehzawad

examples/basic/hello_world.py
examples/basic/hello_world_gpt_5.py
examples/basic/hello_world_gpt_oss.py
examples/basic/stream_text.py
examples/basic/stream_items.py
examples/basic/stream_function_call_args.py
examples/basic/dynamic_system_prompt.py
examples/basic/prompt_template.py
examples/basic/agent_lifecycle_example.py
examples/basic/lifecycle_example.py

ehzawad

# PyTorch 2.x; Apple Silicon; runs on MPS if available, else CPU.
import math
import torch

if torch.backends.mps.is_available():  # official API to detect MPS
    device = torch.device("mps")
else:
    device = torch.device("cpu")

# --- Quantization helpers (per-tensor, symmetric) ---

ehzawad

import React, { useMemo, useRef, useState, useEffect } from "react";

// Simple FAISS-like toy: IndexFlatL2
class IndexFlatL2 {
  constructor(d) {
    this.d = d; // dimensionality
    this.vectors = []; // each item: {id, x: number[], label?: string}
    this._nextId = 1;
  }
  add(vecs) {

ehzawad

Looking at your GeForce RTX 2050 laptop GPU, here's what I recommend:

Best Driver Options for RTX 2050 Laptop:

1. nvidia-driver-575-open (RECOMMENDED)

• Version: 575.64.03
• Why: Ubuntu specifically recommends this for your RTX 2050
• Benefits for laptops: the driver also added NVIDIA Dynamic Boost support for Linux notebooks running on battery power
• CUDA: Fully compatible with CUDA 12.9

ehzawad

#!/usr/bin/env python3
"""
Run GPT-OSS using the Metal implementation for Apple Silicon
This requires downloading and converting the model first.
"""

import os
import subprocess
import sys