Predicting Model Performance › Prediction Reliability
Language AI HandbookAll chapters Predicting Model Performance › Prediction ReliabilityOne section · Sections in this chapter5 / 10 ↓
Reading tools Definitions & inline links
Choose your expertise level to adjust how many terms are explained. Beginners see more tooltips, experts see fewer to maintain reading flow. Hover over underlined terms for instant definitions.
Article links
Make inline references clickable
When can we trust scaling law predictions? Several conditions affect reliability. Understanding these conditions helps practitioners know when to rely heavily on predictions and when to treat them as rough guides requiring empirical verification.
Conditions for Reliable Prediction
Scaling law predictions are most reliable when:
1. Training conditions remain constant. The scaling laws assume consistent training procedures, including similar architectures, hyperparameters, data distributions, and optimization strategies. Changes to any of these can shift the scaling curves. A scaling law fit on GPT-style autoregressive models won't accurately predict T5-style encoder-decoder models. Similarly, changing from AdamW to a different optimizer, or from standard attention to flash attention, can introduce systematic offsets. This doesn't make predictions useless; the slopes often remain similar, but the intercepts can shift.
2. Extrapolation is limited. Predictions within 10× of training data are generally reliable. Predictions beyond 100× are speculative. The Chinchilla paper fit on models up to 16B parameters, yet predictions for 70B models proved reasonably accurate, though this represented only ~5× extrapolation. A rule of thumb is to treat the uncertainty as roughly doubling for each order of magnitude of extrapolation beyond the training regime.
3. The capability doesn't exhibit phase transitions. Smoothly improving capabilities like perplexity and HellaSwag accuracy extrapolate better than emergent capabilities like chain-of-thought reasoning. Before attempting capability prediction, examine whether existing data shows smooth trends or hints of nonlinear behavior. Smooth trends at small scale suggest (but don't guarantee) smooth behavior at large scale.
4. Data quality is maintained. Scaling laws fit on high-quality data don't transfer to low-quality data. As training datasets grow, maintaining quality becomes increasingly difficult. This creates a source of systematic prediction error. If you're fitting on carefully curated data but will train the final model on noisier data due to volume requirements, predictions will be systematically optimistic.
Historical Prediction Accuracy
The track record of scaling law predictions provides calibration for future forecasts. Examining cases where predictions succeeded and failed helps us understand reliability in practice:
Historical scaling law prediction accuracy: ---------------------------------------------------------------------- Model Predicted Actual Error Extrap ---------------------------------------------------------------------- GPT-3 (175B) 2.65 2.72 -2.6% 12× Chinchilla (70B) 2.10 2.09 +0.5% 4× LLaMA-2 70B 1.85 1.87 -1.1% 8×
These examples show that predictions within ~5% are achievable for moderate extrapolation factors. The Chinchilla prediction was remarkably accurate at only 4× extrapolation. Even GPT-3, with 12× extrapolation, came within 3% of prediction. However, this historical accuracy reflects models trained under similar conditions to the scaling law fits. Novel architectures or training procedures can produce larger deviations. The track record supports confidence but not overconfidence.
Known Failure Modes
Scaling law predictions fail systematically in several scenarios:
Architecture discontinuities. The transition from attention mechanisms to linear attention, or from dense to sparse (MoE) models, introduces new scaling relationships. Predictions fit on dense transformers don't transfer to MoE models without refitting. MoE models have different effective parameter counts and different compute-performance relationships. Each architecture requires its own scaling study.
Data exhaustion. When models train on all available high-quality data and must use lower-quality sources, performance degrades faster than loss-based scaling laws predict. The scaling law assumes infinite high-quality data, but reality imposes constraints that scaling laws don't capture. As training corpora grow larger, data quality and diversity become binding constraints that idealized scaling laws don't capture.
Capability phase transitions. Some capabilities emerge suddenly, as we'll examine in Part XXX. No amount of loss extrapolation predicts these transitions from first principles. If a capability hasn't appeared at any scale you've tested, loss trends provide no guidance about when it might emerge.
Optimization instabilities. Larger models encounter training instabilities that smaller models don't. Loss spikes and training divergence aren't captured in smooth scaling laws, which assume successful training. A prediction that 100B parameters will achieve loss 1.9 doesn't account for the possibility that training at that scale proves unstable and requires restarts, different hyperparameters, or architectural modifications.