mercor

what is human data

raw material that teaches models how to think, reason, act

how is text broken down?

tokens

what are tokens

numerical chunks that models process mathematically

how does a model predict what will come next?

assigns probability to every possible next token

what is the probability scheme of models?

stochastic, draws from a probability distribution

what are the two phases of training

pre training, post training

what is pre training

trained on massive internet text

what is the objective and result of pre training

predict the next token, create vast knowledge and fluency

what is the post training

trained on targeted, human-designed signals

what is the objective and result of post training

answer this, prefer this, meet this criteria, then create a genuine assistant

what is a reward signal

a way to grade an output, eval

what is an eval

test for an AI system, give input, apply grading logic to measure success

what is the eval lifecycle

task, ai/agent, transcript, grader, score

what is a single run of 1 task

trial

what are the four ways to shape model behavior

1. supervised fine-tuning

2. preference labeling / RHLF

3. rubric-based evaluation

   1. reinforcement learning environments

what is supervised fine-tuning

giving a student practice problems with correct answers — learning by what is right

what is preference labeling and RHLF

asking a teacher which of 2 essays is better, learning from comparison

what is rubric-based evaluation

giving a teacher a grading guide instead of a model answer, write criteria once, auto-grade thousands of responses

what is reinforcement learning environments

like a flight simulator for pilots — put in realistic environment and see if they can fly the plane

when do you use SFT

when there are clear, objective criteria (math, coding)

when do you use preference labeling?

when ‘better’ is hard to articulate (writing)

when do you use rubrics?

use when you can grade things on criteria

when do you use RL environments

when the model must take real actions (navigate website, book a flight)

what are the failures of SFT

wrong format, structure, steps

what are the failures of RL from human feedback?

capable but wrong tone, style, character

what are the failures of rubrics?

inconsistent quality on complex, subjective reasoning

what are the failures of RL environments

cannot interact with tools and systems beyond the chat box

which of the four methods requires no grader?

supervised fine-tuning

what kind of eval is preference?

automatic evaluator

what kind of eval is rubrics?

same artifact grades and trains

what kind of eval is RL environment?

an eval of end-state

what is a single-turn?

one prompt, one response, one grade

what is multi-turn?

back and forth conversation

what is an agent?

using tools across many turns

how does this affect a projct?

the complexity of a project sets the scope, cost and staffing

what is pass@k?

capability, probability of success in k tries, increases with more attempts

what is pas^k

reliability, probability all k trials succeed, goes down with more k

what is the hill climbing loop?

train, measure, find weaknesses, generate new ideas, repeat

APEX (Mercor)

AI productivity index, can frontier models do high-value knowledge work, tests economically relevant capability

GDPval (openAI)

real-world, economically valuable tasks across 44 occupations in the 9 largest US GDP sectors

SWE-bench verified (openAI)

resolve real github issues — given a repo and issue, generate a patch that passes the test suite

humanity’s last exam (scaleAI)

a massive, extremely hard benchmark probing whether models surpass human-level on high-stakes tasks

HealthBench (openAI)

medical questions, diagnosis, and healthcare reasoning grounded in clinical relevance and safety

tau-bench (multi-turn)

simulated conversational agents in retail and airline settings - a second LLM plays the user