Data Shapley in One Training Run: Valuing Every Training Sample from a Single Pass

Forum topic · 小凯 · 2026-05-03

Summary

This forum post discusses the paper 'Data Shapley in One Training Run' (May 2026), which addresses the data valuation problem in RLHF and large language model fine-tuning. Traditional Shapley value estimation requires retraining a model tens of thousands of times, making it computationally prohibitive. The proposed method instead extracts each training sample's contribution from a single full training run by monitoring gradient dynamics throughout training: it records the influence each data point exerts on the loss function at different model states along the training trajectory, then integrates these quantities over the training timeline—an influence-function-style approach. The result is a per-sample contribution score that can identify harmful samples (e.g., biased or low-quality instructions in an RLHF dataset) and high-value samples, even inside billion-parameter black-box models. The author frames this as replacing manual data auditing with a mathematically grounded 'data valuation engine,' arguing that the next phase of LLM competition is data extraction and curation rather than raw volume. Key takeaway: without knowing which training samples help or harm, compute spend on massive datasets is wasted; per-sample valuation enables training stronger models with a fraction of the data.

Data Shapley in One Training Run: Are You Feeding Your Model Everything—Or Hunting the Poisoned Mole?

After reading Data Shapley in One Training Run (2026.05)—a paper tackling the data valuation problem in RLHF and large-model fine-tuning—it feels like data annotators finally have an absolute physical yardstick for their KPIs.

To explain why today's AI models are so often derailed by one or two dirty data points, let's talk about slicing the cake.

1. The Status Quo: The Mole Slacking Off Inside Your Training Set

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

The pain point: You spend $1M hiring people to write 100K high-quality instructions. But 500 samples with discriminatory bias or broken logic slip into the mix. After training, the model got dumber—and you have no idea which samples poisoned it. Inside a hundred-billion-parameter black box, all the data is stirred into one pot. This is the physical untraceability of training gradients.

2. Data Shapley in One Training Run: The Accountant with a Built-In Microscope

Economics has a concept called the Shapley value, which can fairly compute each person's contribution to a team. In deep learning, however, computing it traditionally required retraining the model tens of thousands of times—compute costs that could bankrupt you.

The authors of this paper took a bold approach: skip the tens of thousands of retrains—use calculus to extract each data point's true value from a single complete training run.

Physical intuition (integrating the gradient trajectory): Instead of looking only at the final model, it watches the gradient dynamics each sample produces throughout training. It records: when the model is in state A, how much "push" does this sample give the loss function? When the model is in state B, how much push then?
Elevated influence functions: By precisely integrating these microscopic quantities over the entire training timeline, the system directly outputs a most-wanted list: sample #8472 is garbage—it contributed -5.4 to the model's intelligence; sample #1024 is a masterpiece—it contributed +12.8.

3. A Feynman-Style Judgment: Quality Is Precisely Peeling Apart Microscopic Gradients

If "data cleaning" relies on manual spot checks, it's slash-and-burn farming.

Real science is being able to use a mathematical ruler, amid the chaos of backpropagation, to cleanly separate the torque every sample exerts on the model's evolution.

One-run Data Shapley tells us: the second half of the LLM war is a war of data extraction.

When engineers can wield a scalpel—precisely excising every piece of rotten meat from the training set and amplifying the weight of core stem-cell samples—we can train models far surpassing today's with one-tenth of the data.

Takeaway:

When preparing 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 millions of GPU dollars you burn are just paying for a blind statistical game.