Activation and trajectory analysis
Probe selected hidden-state dimensions at a layer, compare Base and SFT activations, and trace a dimension across the model's full layer stack.
Research environment
ModelWalker is a live research environment for studying how fine-tuning changes large language model behavior. A browser interface and an MCP server share one GPU-resident model, allowing the same OLMo-3 Base and SFT experiments to be run interactively or from an agent workflow.
What it makes observable
Causal ablation
Zero a specific dimension at a decoder block and measure its effect on next-token operator mass against a control dimension.
Token-level model comparison
Inspect raw logits, probabilities, vocabulary ranks, and generation traces to understand how model behavior changes after fine-tuning.
Shared-model research workflow
The web GUI and MCP server route requests to a single backend that manages model residency and caching. This keeps experiments reproducible while avoiding redundant model loads, so an investigation can move naturally between visual analysis and agent-led probing.