Data Shapley in One Training Run: Precise Data Valuation for LLM Fine-Tuning

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. Traditionally, computing Shapley values for training data requires retraining the model tens of thousands of times, making it computationally prohibitive. The paper's method instead extracts each sample's contribution from a single training run by tracking gradient dynamics throughout the training trajectory. It records how much each data point pushes the loss function at different model states, then integrates these quantities over the training timeline to produce per-sample contribution scores. The result is a precise 'wanted list' that can identify harmful samples (e.g., biased or noisy data poisoning a fine-tuned model) as well as highly valuable ones, solving the problem that individual toxic samples are usually impossible to trace inside a black-box model. The author argues that the next phase of the LLM race will be a war over data refinement: rather than chasing massive dataset scale, engineers should deploy data valuation engines to surgically remove bad samples and amplify high-quality ones, achieving better models with a fraction of the data and GPU cost.

Data Shapley in One Training Run: Feeding Data "Big-Pot Meals" or Precisely Finding the Poisoning Mole?

After reading Data Shapley in One Training Run (2026.05) — a remarkable 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 "KPI reviews".

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

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

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

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

2. One-Run Data Shapley: An Accountant with a Built-in Microscope

Economics has a concept called the Shapley value, which can compute each person's contribution to a team with perfect fairness. In deep learning, however, computing it previously required retraining the model tens of thousands of times — a bankruptingly expensive process.

The paper's authors took a bold approach: no tens of thousands of retrains — just one complete training run, using calculus to squeeze out the "gold content" of every single sample.

Physical picture (integrating the gradient trajectory): It doesn't look at the final model. Instead, it watches the gradient dynamics generated by each data point 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?
Influence functions, upgraded: By precisely integrating these microscopic quantities across the entire training timeline, the system directly outputs a "wanted poster": sample #8472 is garbage, contributing -5.4 to model intelligence; sample #1024 is a masterpiece, contributing +12.8.

3. A Feynman-Style Judgment: Quality Is "Precise Dissection of Microscopic Gradients"

So-called "data cleaning" done by manual spot-checks is slash-and-burn farming. Real science is being able to use a mathematical ruler to strip out, from the chaotic churn of backpropagation, the exact torque each 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 refinement.

When engineers can wield a scalpel to precisely excise every piece of "rotten flesh" from a training set and multiply the weight of the core "stem cells," we can train models far superior to their predecessors with one-tenth of the data.

Takeaway:

When preparing large-scale training data, stop blindly chasing "tens of millions" of samples. Turn on 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 you burn on GPUs are just paying for a blind statistical gamble.

---

*Note: This is an English translation of a Chinese forum post. The original includes a review-style commentary on the referenced paper.*