oneryalcin

Fine-Tuning Sparse Encoders for Neural Sparse Retrieval: Complete Reproducible Guide (SPLADE, OpenSearch Neural Sparse, CSR)

Fine-Tuning Sparse Encoders for Neural Sparse Retrieval

A Complete Reproducible Guide (SPLADE, OpenSearch Neural Sparse, CSR)

Last updated: 2026-02 Scope: Covers SPLADE / SPLADE++ / SPLADE-v3, OpenSearch Neural Sparse v1–v3 (including inference-free), and CSR (Contrastive Sparse Representation). Includes full training recipes, loss functions, architecture internals, and practical tips.

---

oneryalcin

#!/usr/bin/env python3
"""Extract Claude Code session JSONL to lean readable markdown.

Usage: python3 claude_session_extract.py <session.jsonl> [output.md]

If output.md is omitted, writes to <session_id>.md in current directory.
"""
import json
import sys
from pathlib import Path

oneryalcin

pen2pptx

Convert .pen design files to PowerPoint presentations.

Installation

npm install pptxgenjs

oneryalcin

#!/usr/bin/env python3
"""
RLM v2 (Recursive Language Model) Module for Smolagents

Implements DSPy RLM-style strategies for handling large contexts:
- Peeking: Look at data structure before processing
- Grepping: Use string/regex matching before LLM calls
- Partition + Map: Chunk data and process with batched sub-LLM calls
- Summarization: Hierarchical compression for understanding the whole
- Hybrid: Combine strategies as needed

oneryalcin

#!/bin/bash
# Crawl and download Claude platform docs as markdown

set -e

BASE_URL="https://platform.claude.com/docs/en"
SITEMAP_URL="https://platform.claude.com/sitemap.xml"
WORK_DIR="/tmp/claude-docs"
OUT_DIR="$WORK_DIR/docs"
URLS_FILE="$WORK_DIR/urls.txt"

oneryalcin

"""
SDK Patch with Subagent Support - Adds .raw field to SDK message types with maximum fidelity.

Provides drop-in replacements:
- ClaudeSDKClientWithRaw: replaces ClaudeSDKClient
- query_with_raw: replaces query

Data sources for .raw:
- user/assistant: JSONL file (has parentUuid, timestamp, isSidechain, etc.)
- result: Raw CLI output (has modelUsage, errors, permission_denials)

oneryalcin

Luxical: The Engineering Deep Dive

"Transformers without the Heavy Lifting"

Authorship Note: This document was compiled during an interactive exploration session simulating a "Feynman Lab" environment. It deconstructs the Luxical project to explain how modern engineering (Rust, Numba, Distillation) allows simple arithmetic to achieve state-of-the-art results.

---

Table of Contents

1. The Problem: The Efficiency Gap

oneryalcin

Network Latency Analysis Report

Call Center Connectivity to AWS US-East-2 (Ohio)

Monitoring Period: November 4-9, 2025

---

Executive Summary

Call center users in Pristina (Kosovo) and Diber (North Macedonia) reported app quality degradation on Saturday, November 8, 2025 at 6:00 PM UTC, with brief improvement around 6:30 PM, followed by recurring issues around 8:00 PM. This report analyzes network latency data collected from RIPE Atlas probes monitoring both call center network paths, identifying the specific network segments causing degradation.

oneryalcin

RIPE Atlas Call Center Monitoring Documentation

Executive Summary

Monitoring setup to detect evening latency spikes affecting call centers in Pristina (Kosovo) and Diber (North Macedonia) connecting to AWS US-East-1 (Virginia) and US-East-2 (Ohio) regions.

1. Call Center Infrastructure

Call Center Locations and IPs

Pristina (Kosovo):

oneryalcin

Skills Pattern: Just-In-Time Context Injection for LLM Agents

Right. Focus on injection mechanics and why it works, not what skills exist.

Core Mechanics

1. Just-In-Time Context Injection

Not "here's all documentation upfront" - it's pull, not push: