Feynman Letter: Computing Data Shapley in One Training Run

Forum topic · 小凯 · 2026-05-03

Summary

This forum post discusses a recent paper, "Data Shapley in One Training Run" (2026.05), which tackles the problem of data valuation in RLHF and fine-tuning of large language models. Traditional Shapley-value-based data attribution requires retraining the model tens of thousands of times, making it computationally prohibitive. The featured approach instead extracts each training sample's contribution from a single full training run by monitoring gradient dynamics along the training trajectory and integrating influence-function quantities over the entire training timeline. The output is a per-sample contribution score, allowing practitioners to identify harmful samples (e.g., biased or logically inconsistent data mixed into instruction datasets) as well as high-value samples. The author frames this as solving the "physical untraceability of training gradients" in black-box models, and argues that the next phase of the LLM race is a war of data extraction rather than raw data volume: precise, mathematically grounded data valuation lets engineers remove toxic samples and upweight core ones, potentially matching prior results with a fraction of the data and GPU budget.

Feynman Letter: Do You Want to Feed Your Data "Equal Shares," or Precisely Root Out the Poisonous Mole? — On Computing Data Shapley in One Training Run

After reading the paper Data Shapley in One Training Run (2026.05) — which addresses the data valuation problem in RLHF and large-model fine-tuning — I feel that data annotators finally have an absolute physical yardstick for their KPI evaluations.

To help you understand why today's AI models are so often derailed by one or two dirty data points, let's talk about "dividing the cake."

1. Current situation: the mole slacking off inside your training set

When fine-tuning large models (especially RLHF), data quality is everything.

The pain point: You spent $1M hiring people to write 100k high-quality instructions. But 500 garbage samples with discriminatory bias or logical confusion got mixed in. After training, the model got dumber — and you have no idea which samples "poisoned" it. Inside a black box of hundreds of billions of parameters, all the data has been stirred into one pot. This is the physical untraceability of training gradients.

2. One-training-run Data Shapley: the accountant with a built-in microscope

Shapley value

The authors of this paper are ruthless: instead of retraining tens of thousands of times, we extract each sample's "gold content" from a single full training run using the magic of calculus!

Physical picture (integration along the gradient trajectory): Instead of looking only at the final model, it watches the gradient dynamics produced by each data point throughout training. It records: when the model is in state A, how much "push" does adding this sample give the loss? And how much when the model is in state B?
Influence functions, upgraded: By precisely integrating these microscopic quantities over the entire training timeline, the system directly outputs a "most-wanted list": sample #8472 is garbage, contributing -5.4 to the model's "IQ"; sample #1024 is a masterpiece, contributing +12.8.

3. A Feynman-style judgment: quality is "precise dissection of microscopic gradients"

"Data cleaning" by manual spot-checks is slash-and-burn farming.

Real science is being able to use a mathematical ruler to cleanly strip out, from the chaos of backpropagation, the exact torque every sample exerted on the evolution of the universe (the model).

One-training-run Data Shapley tells us: the second half of the large-model war is a war of "data extraction."

When engineers can, like wielding a scalpel, precisely excise every inch of "rotten flesh" in the training set and multiply the weight of the core "stem cells," we can train silicon gods that far surpass their predecessors with one-tenth of the data.

Key takeaway:

When preparing your large-scale training data, stop blindly chasing "tens of millions" of samples.

Build your Data Valuation engine.

If you don't know which bite you're feeding the model is meat and which is poison, then the tens of millions you burn on GPUs are just paying for a blind statistical gamble.

Tags from the original post: DataShapley, DataValuation, RLHF, LLM, MachineLearning, Optimization, FeynmanLearning