Applied Models

Applied Models is a public home for empirical science and applied research on generative models.

The most important questions about generative models are still open. This is the work of finding out — direction by direction, experiment by experiment.

Research Directions

Fundamentals How do architecture choices and training dynamics produce the behaviors we observe at scale? Starting from model internals, not from benchmark scores.

Interpretability Which circuits drive specific model behaviors, and can they be precisely identified and edited? Mechanistic analysis of what models actually learn, not just what they output.

Alignment Science When does alignment hold under distribution shift and pressure, and how does it degrade? Behavioral consistency studied empirically across contexts.

Post Training What does each training stage — fine-tuning, DPO, distillation — actually change inside the model? Behavior, or internal representations?

RL Does RL genuinely improve reasoning, or does it optimize for the appearance of reasoning? RLHF, GRPO, and process reward models studied from outcome to mechanism.

Evals What would a rigorous, falsifiable eval of model capability actually require? Hypothesis-driven suites for capabilities, failure modes, and behavioral consistency. Starting with Gemma.

Industry & Enterprise Where do retrieval-augmented generation, structured output, and constrained decoding reliably break — and what can be done about it?

Why These Questions

Genuine unknowns Every experiment starts with something that doesn't have a clear answer yet — not a demonstration of what works.

Narrow enough to test One model, one configuration, one measurable outcome. Small scope makes real answers possible.

Empirical first What the model actually does under specific conditions — not what theory predicts.

Compounding record Seven directions, accumulating real evidence over time. The record stays public.

What Gets Published

Traceable results Every finding links back to a run done here.

Reproducible method Clear enough that someone else could replicate it with the same setup.

Honest failure Inconclusive runs and broken setups documented, not hidden. The record is not a highlight reel.

No repackaging Every claim traces to a run done here, not to someone else's paper.

Reproducibility Standard

Every released artifact should be reproducible by someone else. That means:

• a clear method with enough detail to replicate
• evaluation criteria set before running
• failure modes and limitations documented
• code, configs, or prompts made available

If it can't be reproduced, it's a note — not a release.

Behind the Work

Applied Models is a personal initiative by the builder behind Prachalabs.com and Pracha.me.

Models and Datasets

Public model assets: appliedomodels on Hugging Face

Publication Boundary

Every release comes from direct implementation and first-hand observation.