Instead of mixing AI training data by where it came from, what if you sorted it by the skill it teaches?
Can training data organization by capability outperform source or task-based mixing?
This explores whether grouping or selecting training data by the skill it builds beats the usual approach of blending by where data came from (web, code, books) or by task label.
This explores whether grouping or selecting training data by the skill it builds beats the usual approach of blending by source or task label. No note in the collection runs that head-to-head comparison. But several attack it from different angles, and they lean toward yes, with a catch: a piece of data's "capability" depends on the model receiving it.
The most direct evidence is LESS. It ignores source and task labels and scores each training example by whether its gradient looks like the target skill's. Training on the selected 5% beats training on the full dataset, because mixed datasets contain examples that actively hinder a specific skill by pulling the model's reasoning strategy away from what the task needs (Can we train better models on less data?). So a source-based blend isn't neutral, since parts of it can do harm. StructTuning makes a similar point from the other side. It organizes chunks into an auto-generated taxonomy of concepts and reaches about half of full-corpus performance with 0.3% of the data. The model learns where a fact sits in a structure, the way a student learns from a textbook, rather than absorbing raw text (Can organizing knowledge structures beat raw training data volume?).
Ordering by capability type also matters. Omni-Thinker found that structured domains push a model's output entropy down while creative domains push it up. Training the structured tasks first beat joint training, which is the blended approach, by 6.2%, because it stopped entropy collapse from damaging open-ended skills (Does training order reshape how models handle different task types?). Transformer2 takes the same instinct to the weights. Instead of blending everything into one model, it tunes a small expert vector per skill and mixes them at inference, beating LoRA with fewer parameters (Can models dynamically activate expert skills at inference time?).
Task labels and source mixes look like weak organizing principles for two reasons. Models tuned on semantically empty or deliberately wrong instructions score about the same as models tuned on correct ones (43% vs a 42.6% random baseline). That suggests what transfers is the output format, not task understanding, so a task label may track the wrong thing (Does instruction tuning teach task understanding or output format?). And a carefully balanced mix may not stay balanced. RL training amplifies one dominant pretraining format within the first epoch and suppresses the rest (Does RL training collapse format diversity in pretrained models?).
The catch is that capability isn't a fixed label on the data. Teacher-refined data can hurt when it sits beyond the student's learning frontier, even when it is objectively better (Does teacher-refined data always improve student model performance?). The reusable unit of reasoning data turns out to be a feedback interface tangled up with the verifier, base model, optimizer and budget. Change any of them and the same data behaves differently (What is the actual reusable unit of reasoning data?). So the approach that wins seems to be organizing data by measuring what the model needs, as LESS does, rather than assigning a capability tag in advance. The collection has no controlled test against source mixing itself. Each result above is measured against a different baseline, such as full data or joint training.
Sources 8 notes
LESS uses low-rank gradient features to select instruction data most similar to target capabilities, and training on the selected 5% consistently outperforms full dataset training. The improvement occurs because mixed datasets contain examples that actively hinder specific skills by shifting reasoning strategy away from task requirements.
StructTuning achieves 50% of full-corpus performance using only 0.3% of training data by organizing chunks into auto-generated domain taxonomies. The model learns knowledge position within conceptual structures rather than raw text patterns, matching how students learn from textbooks.
Omni-Thinker shows structured domains decrease output entropy while creative domains increase it. BWT-guided scheduling—training structured tasks first—yields 6.2% gains over joint training by preventing entropy collapse from damaging open-ended capabilities.
Transformer2 demonstrates that tuning only singular values within weight matrices produces composable expert vectors that dynamically mix at inference without interference, outperforming LoRA with fewer parameters and enabling continual specialization.
Models trained on semantically empty or deliberately incorrect instructions achieve comparable performance to those trained on full correct instructions, achieving 43% vs random baseline 42.6%. The semantic content of instructions appears largely irrelevant; what transfers is knowledge of the output space.
Show all 8 sources
Controlled experiments show RL consistently amplifies one format distribution from pretraining within the first epoch while collapsing alternatives. The winning format depends on model scale, not necessarily performance, and is largely hidden when starting from proprietary pretrained models.
Teacher-refined data degrades performance when it exceeds the student's learning frontier, even if objectively higher quality. Students should filter refinements using their own statistical profile to retain only compatible improvements.
The reusable unit in post-training is a feedback interface entangled with six factors: verifier, base model, lineage, optimizer, scaffold, and budget. Changing any one alters the same data's effect, making attribution tractable only when these are jointly released.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Exploring Format Consistency for Instruction Tuning
- Planted in Pretraining, Swayed by Finetuning: A Case Study on the Origins of Cognitive Biases in LLMs
- A Primer in Post-Training Reasoning Data: What We Know About How It Works
- LESS: Selecting Influential Data for Targeted Instruction Tuning
- Do Models Really Learn to Follow Instructions? An Empirical Study of Instruction Tuning
- Echo Chamber: RL Post-training Amplifies Behaviors Learned in Pretraining
- A Survey on Post-training of Large Language Models
- On the Interplay of Pre-Training, Mid-Training, and RL on Reasoning Language Models