DEV Community

Measuring ML systems honestly Series' Articles

Back to Raihan's Series
How Accurate Is Your Model Right Now? Estimating Accuracy Without Labels
Cover image for How Accurate Is Your Model Right Now? Estimating Accuracy Without Labels

How Accurate Is Your Model Right Now? Estimating Accuracy Without Labels

1
Comments 7
4 min read
AWQ Looked 10 Points Better on GSM8K Until I Stopped Truncating the Answers
Cover image for AWQ Looked 10 Points Better on GSM8K Until I Stopped Truncating the Answers

AWQ Looked 10 Points Better on GSM8K Until I Stopped Truncating the Answers

Comments 4
12 min read
Do LLMs Invent Japanese Law Articles? A Bilingual Benchmark
Cover image for Do LLMs Invent Japanese Law Articles? A Bilingual Benchmark

Do LLMs Invent Japanese Law Articles? A Bilingual Benchmark

Comments
4 min read
Small LLM Judges Approved 11% and 41% of Wrong Answers. Then I Fixed My Own Pairwise Test.
Cover image for Small LLM Judges Approved 11% and 41% of Wrong Answers. Then I Fixed My Own Pairwise Test.

Small LLM Judges Approved 11% and 41% of Wrong Answers. Then I Fixed My Own Pairwise Test.

3
Comments 6
5 min read
GraphProof-QA: Teaching a Small Model to Prove Its Answers
Cover image for GraphProof-QA: Teaching a Small Model to Prove Its Answers

GraphProof-QA: Teaching a Small Model to Prove Its Answers

1
Comments 2
4 min read
Mentioned Is Not Binding: Which Solicitation Clauses Actually Bind the Contract
Cover image for Mentioned Is Not Binding: Which Solicitation Clauses Actually Bind the Contract

Mentioned Is Not Binding: Which Solicitation Clauses Actually Bind the Contract

Comments
3 min read
A Grammar Stopped My Model Inventing FAR Clauses. It Also Made It Refuse the Real Ones
Cover image for A Grammar Stopped My Model Inventing FAR Clauses. It Also Made It Refuse the Real Ones

A Grammar Stopped My Model Inventing FAR Clauses. It Also Made It Refuse the Real Ones

Comments 2
4 min read
I Distilled a 568M Multilingual Model Into a 37M Japanese-English Encoder — Here's What Survived
Cover image for I Distilled a 568M Multilingual Model Into a 37M Japanese-English Encoder — Here's What Survived

I Distilled a 568M Multilingual Model Into a 37M Japanese-English Encoder — Here's What Survived

1
Comments
4 min read
Teaching a Computer to Hear Japanese Pitch Accent
Cover image for Teaching a Computer to Hear Japanese Pitch Accent

Teaching a Computer to Hear Japanese Pitch Accent

1
Comments
2 min read
My Triton Kernel Ties bitsandbytes — and Both Lose to Math
Cover image for My Triton Kernel Ties bitsandbytes — and Both Lose to Math

My Triton Kernel Ties bitsandbytes — and Both Lose to Math

Comments
4 min read
Was My 2,103-Token Tokenizer a Mistake?
Cover image for Was My 2,103-Token Tokenizer a Mistake?

Was My 2,103-Token Tokenizer a Mistake?

Comments
3 min read
DemoDoctor: Bad Robot Demonstrations Are Detectable. Do They Matter?
Cover image for DemoDoctor: Bad Robot Demonstrations Are Detectable. Do They Matter?

DemoDoctor: Bad Robot Demonstrations Are Detectable. Do They Matter?

1
Comments 1
3 min read
Invoice-Check JP: A 3B Model Reads the Invoice, Checks Decide Whether to Trust It
Cover image for Invoice-Check JP: A 3B Model Reads the Invoice, Checks Decide Whether to Trust It

Invoice-Check JP: A 3B Model Reads the Invoice, Checks Decide Whether to Trust It

Comments
4 min read
PO Matching Looked Like It Halved Unsafe Invoice Approvals Until I Checked Why
Cover image for PO Matching Looked Like It Halved Unsafe Invoice Approvals Until I Checked Why

PO Matching Looked Like It Halved Unsafe Invoice Approvals Until I Checked Why

1
Comments 1
5 min read
I Pre-Registered a Four-Way LangGraph Shootout. Plan-and-Execute Cost 2.3x and Bought Nothing I Could Detect
Cover image for I Pre-Registered a Four-Way LangGraph Shootout. Plan-and-Execute Cost 2.3x and Bought Nothing I Could Detect

I Pre-Registered a Four-Way LangGraph Shootout. Plan-and-Execute Cost 2.3x and Bought Nothing I Could Detect

1
Comments 1
6 min read
Edit the Weights, Retrieve the Fact, or Look It Up in a Graph?
Cover image for Edit the Weights, Retrieve the Fact, or Look It Up in a Graph?

Edit the Weights, Retrieve the Fact, or Look It Up in a Graph?

Comments 1
2 min read
Your LLM Judge Is Lying to You, a Little
Cover image for Your LLM Judge Is Lying to You, a Little

Your LLM Judge Is Lying to You, a Little

2
Comments 3
2 min read
The Verb Comes Last: Measuring Why Japanese-English Live Translation Waits
Cover image for The Verb Comes Last: Measuring Why Japanese-English Live Translation Waits

The Verb Comes Last: Measuring Why Japanese-English Live Translation Waits

Comments
3 min read
My Japanese Tutor Agent Works — Except Its Brain Calls Everything a Particle
Cover image for My Japanese Tutor Agent Works — Except Its Brain Calls Everything a Particle

My Japanese Tutor Agent Works — Except Its Brain Calls Everything a Particle

Comments
3 min read
Do Japanese and English Share Neurons? I Trained the Microscope to Check
Cover image for Do Japanese and English Share Neurons? I Trained the Microscope to Check

Do Japanese and English Share Neurons? I Trained the Microscope to Check

Comments
2 min read
One Knob, 19x Latency: What I Learned Benchmarking vLLM on a Single GPU
Cover image for One Knob, 19x Latency: What I Learned Benchmarking vLLM on a Single GPU

One Knob, 19x Latency: What I Learned Benchmarking vLLM on a Single GPU

4
Comments 4
2 min read
Correct, Cited, and Two Years Out of Date: Measuring Version Leakage in RAG
Cover image for Correct, Cited, and Two Years Out of Date: Measuring Version Leakage in RAG

Correct, Cited, and Two Years Out of Date: Measuring Version Leakage in RAG

1
Comments
2 min read
Your RAG Has No Idea What Year It Is: A Frozen Benchmark for Version Leakage

Your RAG Has No Idea What Year It Is: A Frozen Benchmark for Version Leakage

Comments
2 min read
Your MCP Tools Are Leaking: A Benchmark for Description Canonicalisation

Your MCP Tools Are Leaking: A Benchmark for Description Canonicalisation

Comments 1
3 min read