AI 編碼訓練工作實際涉及甚麼？RLHF 程式碼審查樣本解析

There are 231 open AI-training jobs for software engineers on AI Trainer HQ right now, across 10 platforms. The median listed pay is $65/hr, and most roles fall between $30 and $115/hr. Senior roles on 10x Team have a $139/hr median. But "AI coding job" covers very different work, from ranking two Python snippets to debugging an agent across a whole repository. Here's what the work actually involves, which skills pay most, and a realistic sample task.

What RLHF code work involves

RLHF (reinforcement learning from human feedback) for code means engineers produce the judgments a model learns from. Common task types:

• Pairwise ranking: compare two AI solutions to the same problem, pick the better one and explain why (correctness, edge cases, complexity, readability).
• Write the ideal solution: produce a reference answer, often with tests.
• Bug finding: identify why AI-generated code fails and write the fix.
• Prompt writing: design hard, realistic coding problems that current models get wrong.
• Agentic / repo tasks: evaluate an AI agent that edits a real codebase. Does it run, do the tests pass, did it break anything?
• Security review: flag injection, unsafe deserialisation, secrets in code and similar issues.

You don't train models directly. You produce the training data.

Who's hiring and what they pay

From the software engineering hub, data as of 26 October 2026:

Mercor has listed US software engineers at $70–150/hr this month. Several current Mercor roles require you to own a MacBook ($50–60/hr) for macOS app-usage projects. Anyone AI hires developers by country, mostly in Latin America.

Disclosure: AI Trainer HQ earns a referral fee only from micro1. That doesn't affect the data. See /disclosure.

Country access

Coding roles open by country: US 123, India 94, UK 90, Philippines 80, Kenya 78. Coding is one of the most globally open categories, so check yours on the eligibility tool.

Which languages and skills pay most

Listings don't break pay down by programming language, but patterns are clear from the roles:

• Python is the default almost everywhere, so it's table stakes.
• TypeScript/JavaScript, Java, Go, Rust and C++ appear in specialist roles. Rust and C++ often sit in higher-paid systems work.
• Infra and ML engineering (kernels, profiling, LLM serving): one current Mercor MLOps role lists $90–120/hr.
• Language plus coding: Meridial lists "Coding Specialist (Fluent in Japanese / Bengali / Dutch / Swedish / Chinese / Italian / German)" at $65/hr.
• Seniority beats language. 10x Team's $139/hr median is driven by senior engineers and CTO-level reviewers.

A sample task walkthrough

Prompt: "Write a Python function that returns the k most frequent words in a string, case-insensitive, ties broken alphabetically."

Response A:

def top_k(s, k):
    from collections import Counter
    c = Counter(s.lower().split())
    return [w for w, _ in c.most_common(k)]

Response B:

import re
from collections import Counter
def top_k(s, k):
    words = re.findall(r"[a-z']+", s.lower())
    c = Counter(words)
    return sorted(c, key=lambda w: (-c[w], w))[:k]

A strong review:

1. Correctness: A ignores the tie-break rule (most_common keeps insertion order on ties) and treats "word," and "word" as different words. B handles both.
2. Edge cases: B returns fewer than k items if there aren't enough words, which is fine. Note that neither validates k <= 0.
3. Complexity: B sorts all unique words, O(n log n), which is acceptable. Mention heapq.nsmallest for very large inputs.
4. Verdict: B is better. Give a confidence level and a one-paragraph justification that cites the spec.
5. Ideal answer (if requested): B plus input validation and a docstring with an example.

That's the skill being paid for: precise, spec-grounded reasoning, written clearly.

Screen recording (GIF attached: /workspace/aitrainerhq/recordings/aitrainerhq-browse-and-filter-jobs.gif):

How to pass a coding assessment

• Run the code. Many reviewers skip it and miss obvious failures.
• Quote the spec in your justification.
• Don't over-engineer the ideal answer. Graders want correct, idiomatic and readable code.
• Never paste AI output as your answer. Platforms check for it.
• Prepare a short portfolio: a GitHub profile with real work helps a lot on Mercor and 10x Team.
• Full guide: this guide.

Tools and setup that help

• A local environment with Python 3.x, Node and Docker, so you can actually run the code in tasks
• A linter and formatter (ruff, eslint) to spot style issues quickly
• A scratch test file for quick property checks on edge cases
• A reliable second monitor. Long side-by-side reviews are much faster with one.
• For agentic tasks, comfort with git diffs. You'll often be reviewing a patch, not a single function.

Realistic earnings

At the $65/hr median, 10 hours a week is about $650 before tax. Senior engineers on 10x Team can do much better, but those roles have strict screening and fewer seats (10 open right now).

Caveats

• Advertised ≠ earned. Task-based coding work can drop your effective hourly rate if tasks run long.
• Queues dry up. Coding projects end when a model release ships. Keep 2–3 platforms going.
• Confidentiality: you'll often work on code for unreleased models, so don't post tasks publicly.
• A degree isn't always required. Many listings accept equivalent experience, proven through the coding test.

How we got these numbers

Every figure here comes from AI Trainer HQ's own tracking. We pull live jobs only from each platform's own applicant tracking system, re-check them at the source every 24 hours, and convert pay to US dollars per hour. Medians use the midpoint of each role's listed range, and roles that don't state pay are left out. The full method is on our methodology page. Numbers move daily, so treat this post as a dated snapshot and confirm the details on the live listing before you apply.

Next step

Open the software engineering hub, filter by country, and apply to one high-volume platform (Mercor or Meridial) and one senior-focused one (10x Team) if your experience fits.