06 Training and Adaptation 3 min read 686 words

Synthetic Data Generation

In late 2025, the industry has hit the "Data Wall"—the exhaustion of high-quality human text on the web. Synthetic data is now the primary engine for model improvement.

datatrainingalignmentcore
01synthetic shift

After the "Data Wall": The Synthetic Shift

Frontier models (Llama 4, GPT-5.2) are trained on 100T+ tokens. There simply isn't enough human text to sustain this scaling. The 2025 Reality: More than 50% of the training mixture for frontier fine-tuning (and 10% of pretraining) is now synthetic.

SourceHuman DataSynthetic Data
VolumeFixed (Finite)Infinite
QualityVariable (Noisy)Controllable (Purified)
CostHigh (Human Labelers)Cheap (Inference/GPU)
BiasMirror of internetCan be manually balanced
02pattern evolution

Evol-Instruct Pattern (2025 Evolution)

Evol-Instruct is a recursive process where an LLM takes a simple instruction and evolves it into a more complex one.

The Evolution Directions:

  1. Breadth: Increase the number of tasks.
  2. Depth: Add constraints, complicating factors, or multi-step logic.
  3. De-noising: Clean up phrasing to remove "AI-isms."
Pythonpython · 3 lines
123
# Simple Instruction: "Write a function to add two numbers."
# Evolved Instruction: "Write a thread-safe Python class that performs 
# matrix addition with error handling and unit tests, adhering to PEP8."
03feedback rlaif

Constitutional AI & AI Feedback (RLAIF)

Developed by Anthropic and adopted widely in 2025, RLAIF uses a "Constitution" (a set of rules) to guide a model in evaluating and improving its own data.

The Loop:

  1. Propose: Model A generates a response.
  2. Critique: Model B (the constitutional judge) identifies flaws based on guidelines.
  3. Revise: Model A produces a better version based on the critique.
  4. Train: The final (Prompt, Revise) pair is added to the SFT set.
04synthetic data

Verifiable Synthetic Data

The biggest risk of synthetic data is Model Collapse (the model learning its own mistakes). The 2025 Solution: Focus on domains where the "Truth" is verifiable without an LLM.

  • Math: Use Formal Verification (Lean/Isabelle) or Python execution to verify answers.
  • Code: Run generated code against test cases (Unit Tests).
  • RAG: Use "Gold Context" to generate questions where the answer is explicitly in the text.
05diversity

De-biasing and Diversity

Synthetic data is used to "fill the gaps" in human data.

  • Languages: Generating high-quality text in low-resource languages (e.g., Swahili, Marathi) by translating conceptual templates.
  • Logic: Creating 1,000,000 variations of a specific logical fallacy to "harden" the model against it.
06questions

Interview Questions

Q: What is the risk of "Model Collapse" when training on synthetic data?

Strong answer: Model Collapse occurs when a model is trained on data generated by an earlier version of itself. Because the model's distribution is narrower than the real world (it has preferences/biases for certain words and patterns), the training loop becomes a "positive feedback loop" of errors and blandness. By 2025, we mitigate this by:

  1. Mixing in 5-20% "Golden" human-authenticated data.
  2. Using "Verifiable" rewards (Math/Code) so mistakes are never learned.
  3. Using more powerful "Teacher" models to generate data for "Student" models.

Q: How do you ensure the quality of a synthetic dataset of 10 million rows?

Strong answer: We use a Multi-Stage Filtering Pipeline:

  1. Semantic Deduplication: Using embeddings to remove near-identical clusters.
  2. LLM-as-Judge: Sampling 1% of the data and having a stronger model (e.g., GPT-5.2) score it for logic and safety.
  3. Perplexity Filtering: Using a small model to calculate the perplexity of the text. If it's too high (nonsense) or too low (repetitive/simple), it's discarded.
  4. Verifiable Execution: If the data contains code or math, it must pass a local compiler/interpreter check.
07references

References

  • Xu et al. "WizardLM: Empowering Large Language Models to Follow Complex Instructions" (2023)
  • Bai et al. "Constitutional AI: Harmlessness from AI Feedback" (2022)
  • OpenAI. "Weak-to-Strong Generalization" (2023)

Next: Quantization Deep Dive

summary · added by this rebuild

Key takeaways

01

Synthetic data is now the majority ingredient

The page puts over 50% of the frontier fine-tuning mixture and about 10% of pretraining on synthetic text, because high-quality human web text has run out.

02

Evol-Instruct grows instructions along three axes

Breadth adds tasks, depth adds constraints and multi-step logic, de-noising strips AI-isms — turning "add two numbers" into a thread-safe, unit-tested, PEP8 matrix class.

03

Verifiability is the antidote to model collapse

Math checked by Lean or Python, code checked against unit tests, and RAG questions written from gold context keep the model from learning its own mistakes back.

04

Constitutional AI is propose, critique, revise

A generator drafts, a judge model critiques against written guidelines, the generator revises, and only the revised prompt-response pair enters the supervised fine-tuning set.