AI Training Specialist

SKILL.md

# AI Training Specialist

## Overview

This skill provides expert-level AI training services encompassing both evaluation and annotation workstreams. As the AI industry scales, the demand for skilled human annotators and evaluators who can provide high-quality feedback has grown exponentially. This skill bridges the gap between raw human judgment and production-grade AI training data, delivering consistent, calibrated, and actionable outputs across every major task type.

Whether you need RLHF evaluation for a frontier model, multi-turn conversation grading, hallucination detection, or image annotation pipelines, this skill covers the full spectrum of AI training operations. The practitioner brings deep familiarity with the leading platforms, task taxonomies, and quality benchmarks that define modern AI data operations.

---

## Core Competencies

### Evaluation Skills

Evaluation tasks require the practitioner to assess, compare, and refine AI-generated outputs against rigorous quality criteria. These tasks form the backbone of RLHF and constitutional AI pipelines.

#### 1. Response Ranking

Response ranking involves ordering multiple AI-generated responses to the same prompt from best to worst based on criteria such as helpfulness, accuracy, completeness, tone, and safety. The practitioner evaluates pairwise or listwise comparisons, providing transitive-consistent rankings that can be used directly for reward model training.

Key considerations:
- **Preference consistency**: Rankings must be transitive (if A > B and B > C, then A > C) to produce coherent reward signals.
- **Multi-axis judgment**: A response may excel on accuracy but fail on tone; the practitioner must weight axes according to the task specification.
- **Edge case handling**: When responses are nearly equivalent, the practitioner documents the marginal differentiator rather than forcing an arbitrary split.
- **Calibration**: Rankings should reflect meaningful quality differences, not random noise. The practitioner self-calibrates by periodically reviewing past rankings for internal consistency.

#### 2. Response Rating

Unlike ranking (relative), rating assigns absolute scores on a defined scale (e.g., 1-5 or 1-7 Likert). The practitioner applies rubric-defined criteria to each response independently, enabling fine-grained quality measurement.

Key considerations:
- **Rubric adherence**: Every rating must be grounded in the provided rubric. Deviations are documented and flagged.
- **Scale utilization**: The full scale should be used when warranted. Clustering ratings at the extremes or midpoint undermines the data's utility.
- **Inter-rater reliability**: The practitioner maintains awareness of how their ratings compare to population-level distributions, adjusting when systematic bias is detected.

#### 3. RLHF Evaluation

Reinforcement Learning from Human Feedback evaluation is the gold standard for aligning frontier models. The practitioner provides preference judgments that directly train reward models, which in turn guide policy optimization through PPO or similar algorithms.

The practitioner understands the distinction between:
- **Helpfulness evaluation**: Does the response fully address the user's intent?
- **Honesty evaluation**: Is the response factually grounded, or does it fabricate details?
- **Harmlessness evaluation**: Does the response avoid generating unsafe, biased, or harmful content?

Each dimension may require separate evaluation passes, and the practitioner is trained to isolate their judgment on each axis without contamination from the others.

#### 4. Output Editing

Output editing tasks require the practitioner to directly modify AI-generated text to improve quality. This goes beyond rating or ranking—the practitioner produces a corrected version that serves as a demonstration for supervised fine-tuning.

Editing priorities (in order):
1. **Factual accuracy**: Correct any hallucinated or incorrect claims.
2. **Completeness**: Add missing information that a reasonable user would expect.
3. **Clarity and concision**: Remove unnecessary hedging, repetition, or verbosity.
4. **Tone calibration**: Adjust formality, friendliness, or directness to match the target persona.
5. **Safety**: Remove or rewrite any content that violates safety guidelines.

All edits are tracked with change annotations so downstream systems can learn from the diff.

#### 5. Reference Answer Writing

Reference answers serve as ground truth for evaluating model outputs. The practitioner writes authoritative, comprehensive answers to prompts, drawing on domain expertise and research when needed.

Reference answer quality criteria:
- **Completeness**: Address every sub-question and implication in the prompt.
- **Accuracy**: Every factual claim is verifiable. Uncertainty is explicitly stated rather than fabricated.
- **Structure**: Answers are organized with clear headings, logical flow, and appropriate use of lists and tables.
- **Neutrality**: Reference answers avoid editorializing unless the prompt explicitly requests an opinion.

#### 6. Multi-Turn Conversation Evaluation

Evaluating multi-turn conversations requires tracking context across exchanges and assessing coherence, progression, and recovery from errors. The practitioner evaluates entire conversation threads rather than isolated turns.

Evaluation dimensions:
- **Contextual consistency**: Does the assistant remember and correctly reference earlier turns?
- **Error recovery**: When the assistant makes a mistake, does it acknowledge and correct it when the user points it out?
- **Goal progression**: Does the conversation move toward resolving the user's actual need?
- **Turn-level quality**: Each individual response is also assessed for the criteria in single-turn evaluation.

#### 7. Tool-Use Evaluation

Modern AI systems can invoke external tools (search, code execution, API calls). The practitioner evaluates whether tool calls were appropriate, correctly formatted, and whether the assistant correctly interpreted the results.

Evaluation criteria:
- **Tool selection**: Did the assistant choose the right tool for the job?
- **Call correctness**: Were parameters formatted correctly?
- **Result interpretation**: Did the assistant correctly parse and use the tool's output?
- **Necessity**: Was the tool call actually needed, or could the assistant have answered directly?

#### 8. Agent Trajectory Evaluation

Agent evaluation assesses the full trajectory of an autonomous AI agent operating over multiple steps. This includes planning, execution, error handling, and goal achievement.

The practitioner evaluates:
- **Planning quality**: Was the initial plan reasonable given the task?
- **Execution fidelity**: Did the agent follow its plan, and were deviations justified?
- **Error handling**: When steps failed, did the agent recover gracefully?
- **Efficiency**: Did the agent take unnecessary steps or waste tool calls?
- **Goal achievement**: Did the final output satisfy the original objective?

#### 9. Hallucination/Factuality Review

Hallucination detection is among the most critical evaluation tasks for production AI systems. The practitioner identifies factual claims in AI outputs and verifies each one against known information or provided references.

Process:
1. **Claim extraction**: Identify every factual assertion in the output.
2. **Verification**: For each claim, determine if it is supported, contradicted, or unverifiable given the available evidence.
3. **Severity classification**: Minor inaccuracies (wrong dates, approximate numbers) are distinguished from fabrications (invented quotes, non-existent references, fabricated studies).
4. **Reporting**: Produce a structured hallucination report with each flagged claim, its verification status, and the correct information if available.

#### 10. Citation/Grounding Review

For AI systems that cite sources, the practitioner verifies that citations are real, relevant, and accurately represented. This is especially important for retrieval-augmented generation (RAG) systems.

Review checklist:
- Does each citation correspond to a real, accessible source?
- Is the cited source being accurately represented (not quoted out of context)?
- Are key claims supported by citations, or are there unsupported assertions that should have citations?
- Are there redundant or irrelevant citations that pad the response without adding value?

#### 11. Safety/Policy Evaluation

Safety evaluation determines whether AI outputs violate platform-specific or general safety policies. The practitioner applies defined policy frameworks to flag content involving violence, self-harm, illegal activity, sexual content, harassment, or other violation categories.

The practitioner understands:
- **Context dependency**: The same content may be safe in a medical context but unsafe in a casual conversation. Context always matters.
- **Severity gradients**: Not all policy violations are equal. The practitioner categorizes severity (borderline, clear violation, egregious violation) to enable proportional responses.
- **Edge case judgment**: Many safety decisions involve ambiguous or novel situations. The practitioner documents their reasoning for these cases to support policy refinement.

#### 12. Bias/Fairness Evaluation

Bias evaluation examines AI outputs for systematic favoritism or discrimination along protected dimensions (race, gender, religion, nationality, sexual orientation, disability, etc.).

The practitioner looks for:
- **Representational harm**: Does the output reinforce harmful stereotypes?
- **Allocative harm**: Would the output lead to unfair distribution of resources or opportunities?
- **Quality disparity**: Does the output quality vary systematically across demographic groups?
- **Intersectional bias**: Are there bias patterns that only emerge at the intersection of multiple identity dimensions?

---

### Annotation Skills

Annotation tasks involve labeling, categorizing, or structuring raw data to create training datasets. These tasks are the foundation of supervised learning pipelines.

#### 1. Text Classification

Assigning predefined categories to text documents. The practitioner handles multi-label, hierarchical, and fine-grained classification schemes with consistent application of classification criteria.

Best practices:
- Read the full document before classifying; titles and first paragraphs can be misleading.
- When multiple categories apply, apply all relevant labels rather than defaulting to the most salient.
- Flag documents that seem to fall between categories for taxonomy refinement.

#### 2. Sentiment Analysis

Labeling the emotional tone of text, typically on a scale (positive, neutral, negative) or with fine-grained emotion categories. The practitioner distinguishes between the sentiment of the author and the sentiment of subjects mentioned in the text.

Nuances:
- **Sarcasm and irony**: These require contextual understanding beyond keyword matching.
- **Mixed sentiment**: A single text can express both positive and negative sentiments about different aspects.
- **Intensity grading**: Beyond positive/negative, the practitioner assesses the strength of sentiment.

#### 3. Intent Classification

Identifying the user's underlying intent behind a query or statement. Critical for conversational AI, search systems, and customer service automation.

The practitioner maps utterances to intent schemas while handling:
- **Multi-intent queries**: A single message may contain multiple intents.
- **Implicit intents**: Users often state needs indirectly.
- **Out-of-scope intents**: Not every user message fits the predefined intent taxonomy.

#### 4. Summarization Labeling

Assessing the quality of AI-generated summaries along dimensions including faithfulness to the source, coverage of key information, conciseness, and coherence.

Evaluation approach:
- Read the source document first, then the summary.
- Check every claim in the summary against the source. Flag any fabricated or distorted information.
- Assess whether key information from the source is omitted.
- Evaluate whether the summary is appropriately concise without being cryptic.

#### 5. Rewriting Review

Evaluating AI-generated rewrites of text for specific purposes (simplification, formalization, style transfer, etc.). The practitioner assesses whether the rewrite achieves the stated goal while preserving the original meaning.

Key checks:
- **Meaning preservation**: Does the rewrite convey the same information as the original?
- **Goal achievement**: If the task was simplification, is the result actually simpler?
- **Introduction of errors**: Did the rewriting process introduce factual errors or change the author's intent?

#### 6. Translation QA

Quality assurance for machine translation outputs. The practitioner evaluates translations for accuracy, fluency, and cultural appropriateness, comparing source and target texts.

Evaluation dimensions:
- **Adequacy**: Does the translation convey all the information from the source?
- **Fluency**: Does the target text read naturally in the target language?
- **Terminology consistency**: Are domain-specific terms translated consistently?
- **Cultural adaptation**: Are culture-specific references handled appropriately?

#### 7. Search Relevance Rating

Rating the relevance of search results or retrieved documents to a query. This is fundamental to training and evaluating retrieval systems and search engines.

Rating approach:
- Understand the user's information need from the query.
- Rate each result independently before comparing.
- Use the full relevance scale (e.g., Perfect, Excellent, Good, Fair, Bad) rather than binary relevant/irrelevant.
- Consider both topical relevance and utility (does the result actually help the user accomplish their goal?).

#### 8. Content Moderation

Labeling content for policy violations, including hate speech, harassment, spam, self-harm, violence, and sexual content. The practitioner applies platform-specific policies consistently and documents edge cases.

Critical considerations:
- **Cultural context**: What constitutes hate speech or harassment varies across cultures and languages.
- **Nuance and code-switching**: Moderation must account for in-group language use, reclaimed slurs, and contextual appropriation.
- **Freedom of expression balance**: Over-moderation can suppress legitimate speech. The practitioner flags borderline cases for policy team review.

#### 9. OCR/Transcription Correction

Correcting errors in OCR output or automated transcriptions. The practitioner identifies and fixes systematic and random errors, producing clean text that accurately represents the source material.

Common error patterns:
- Character confusion (0/O, 1/l/I, 5/S)
- Missing or inserted characters due to image quality
- Formatting loss (columns, tables, headers) misinterpreted as linear text
- Speaker diarization errors in audio transcription

#### 10. Document Labeling

Assigning structured metadata to documents: topics, entities, relationships, key phrases, and document-level attributes. This supports knowledge base construction and information extraction pipelines.

Labeling approach:
- Follow the provided schema precisely; do not invent new labels without approval.
- Apply entity linking when applicable (linking mentions to canonical entity records).
- Document confidence levels when the schema allows it.

#### 11. Image Classification

Categorizing images according to predefined taxonomies. This includes object recognition, scene classification, and content type labeling.

Quality practices:
- Examine the entire image, not just the most prominent element.
- Apply all relevant labels in multi-label tasks.
- Flag ambiguous or low-quality images rather than forcing a classification.

#### 12. Bounding Box Labeling

Drawing precise bounding boxes around objects of interest in images. This is fundamental for training object detection and localization models.

Best practices:
- Tight fit: Boxes should be as tight as possible while fully containing the object.
- Consistent criteria: Define and consistently apply rules for occluded, truncated, and small objects.
- Attribute labeling: When applicable, attach correct attributes (class, occlusion level, truncation) to each box.

---

## Platform Knowledge

The practitioner maintains deep operational familiarity with the following platforms:

### Scale AI
The market leader in AI data infrastructure. Scale offers the highest-quality evaluation and annotation tasks, often for frontier model developers. Tasks include RLHF evaluation, red-teaming, and complex multi-step annotation. Onboarding typically requires passing qualification assessments. Pay rates are among the highest in the industry for qualified annotators.

### Remotasks
Now integrated with Scale AI's ecosystem, Remotasks provides a broader range of task types with variable complexity. Good entry point for building annotation skills before qualifying for Scale's premium tasks. Tasks include image annotation, text classification, and basic evaluation.

### DataAnnotation.tech
Specializes in AI training tasks with a focus on coding, writing, and evaluation. Known for relatively high pay rates and flexible scheduling. Tasks range from simple classification to complex multi-hour evaluation projects. The platform uses skill-based qualification to gate access to higher-paying tasks.

### Surge AI
A premium data annotation platform that works with top AI labs. Surge emphasizes quality over quantity and typically requires strong demonstrated skills for access. Tasks include RLHF evaluation, prompt engineering, and specialized domain annotation.

### Appen
One of the oldest and largest data annotation platforms. Offers a wide variety of task types but generally at lower pay rates than newer platforms. Good for volume and variety, with tasks ranging from search relevance rating to speech transcription.

### Prolific
Primarily a research participant platform, Prolific is increasingly used for AI evaluation studies. It offers strong participant screening and demographic targeting. Tasks tend to be shorter and more research-oriented than production annotation.

### Invisible Technologies
A premium AI training platform that hires annotators as contractors with relatively high pay. Focuses on complex evaluation tasks including RLHF, red-teaming, and specialized domain evaluation. The application process is selective.

### Toloka
A crowdsourcing platform offering a wide range of microtask annotation. Pay rates vary significantly by geography and task type. Good for high-volume, lower-complexity annotation tasks. Includes quality control mechanisms like golden sets and overlap settings.

---

## Qualifier Strategy

Most AI training platforms gate access to premium tasks behind qualification assessments. The practitioner provides strategic guidance for passing these qualifiers:

### Start with Practice Skills
Before attempting a paid qualifier, practice the underlying skill in free or low-stakes environments:
- Use open-source datasets (HuggingFace, Kaggle) to practice annotation and evaluation.
- Study existing evaluation rubrics from published RLHF datasets (e.g., OpenAssistant, UltraFeedback).
- Build familiarity with the task interface by completing available tutorial tasks.

### Build Judgment Muscle
Qualification is not about raw intelligence—it is about calibrated judgment. Develop this by:
- Practicing consistency: Evaluate the same examples multiple times and check your agreement rate.
- Studying rubrics: Understand exactly what each score level means before rating.
- Seeking feedback: Compare your judgments to published gold standards when available.
- Developing a pre-evaluation checklist: Before rating, quickly assess the response on each evaluation axis.

### Positioning for Non-Developers with Prompt Engineering Experience
You do not need to be a software engineer to excel at AI training. Your competitive advantages include:
- **Prompt engineering intuition**: Understanding how prompts work means you understand how models can fail. This makes you naturally good at evaluation.
- **Communication skills**: Many evaluation tasks require writing explanations for your ratings. Strong writers have an edge.
- **Domain expertise**: Whether your background is in law, medicine, finance, or the humanities, specialized knowledge is extremely valuable for domain-specific evaluation tasks.
- **Critical thinking**: The ability to analyze arguments, identify logical fallacies, and assess evidence quality translates directly to evaluation quality.

---

## Service Offerings and Pricing

### Hourly Rate: $25/hour
Applies to ongoing annotation and evaluation work including:
- RLHF evaluation tasks
- Text annotation and classification
- Response ranking and rating
- Content moderation and review

### Flat Rate: $150 for Full Qualifier Prep
A comprehensive package to prepare you for platform qualification assessments:
- Platform-specific strategy session
- Practice tasks with feedback
- Rubric walkthrough and calibration exercise
- Mock qualifier with detailed feedback
- One follow-up session after the qualifier attempt

### What You Get
- A practitioner with demonstrated skill across all evaluation and annotation task types
- Deep platform knowledge to optimize your workflow and maximize earnings
- Consistent, calibrated judgment that produces high-quality training data
- Clear communication about task requirements, edge cases, and quality concerns

---

## Getting Started

1. **Assessment**: Share your background, experience level, and target platforms.
2. **Skill mapping**: We identify which evaluation and annotation skills align with your strengths.
3. **Training plan**: A customized plan to build the skills you need for your target tasks.
4. **Qualifier preparation**: Focused preparation for specific platform qualification assessments.
5. **Ongoing support**: Continued guidance as you take on more complex tasks and advance to higher-paying opportunities.

The AI training industry is growing rapidly, and skilled evaluators and annotators are in high demand. Whether you are looking for flexible side income or a full-time career in AI data operations, this skill provides the expertise and strategy to succeed.