LoginSign Up
← Back to all sparks
R

rollama

ANALYTICS
Velocity0.0

R package rollama by jbgruber — release notes from GitHub.

github.com/jbgruber/rollama ↗

rollama turns a local-LLM wrapper into an instrument for reproducible annotation

local-llmollamatext-annotationstructured-outputreproducibilityr-package
◆Current state
rollama is an R client for Ollama, aimed at researchers running local models for text annotation and embedding rather than at application developers. Version 0.3.0 adds response caching, logprobs output, batched questions, and a reimplemented structured-outputs path with its own vignette, while syncing against upstream Ollama API changes. The package now covers the full loop a computational social scientist needs: prompt, constrain the output shape, read the model's confidence, and cache the result.
◆Where it's heading
Each release has pushed further from chat toward measurement. Early versions added multi-model querying and dedicated embedding models; 0.2.0 brought make_query() for annotation and multi-server dispatch; 0.2.1 added structured output and custom headers. The 0.3.0 combination of logprobs and caching is the clearest statement of intent — those are features you add for people who need confidence scores and reproducible reruns, not for people building chatbots. Keeping pace with the Ollama API is the recurring maintenance cost.
◆Prediction
Expect the annotation path to keep deepening — likely more tooling around logprob-derived confidence and validation of structured outputs — alongside the routine syncing each Ollama API change forces.

◆Recent moves

  1. 4mo ago

    rollama 0.3.0 adds logprobs, caching and batched queries

    ⚡ SPARK

    The release that completes rollama's turn toward research use: logprobs expose per-token confidence, caching makes reruns reproducible and cheap, several questions can be supplied at once, and structured outputs are reimplemented with a vignette. Together they address the three things annotation work needs from a model — parseable output, a confidence signal, and repeatability.

    View source ↗
  2. 1y ago

    Structured output and custom headers

    Structured output arrives first here, along with custom headers for authenticated Ollama endpoints and configurable output formats. It is the groundwork 0.3.0 later rebuilt and documented properly.

    View source ↗
  3. 1y ago

    make_query() for annotation, multi-server dispatch

    make_query() gives annotation work a purpose-built entry point, and queries can be spread across multiple servers — the first release that treats bulk labelling rather than single conversations as the main use case. Embedding and multi-query performance improve, and the default model moves to llama3.1.

    View source ↗
  4. 2y ago

    rollama 0.1.0

    A small release adding check_model_installed and moving the default model to llama3. Version-number housekeeping more than capability.

    View source ↗
  5. 2y ago

    Dedicated embedding models and multi-model queries

    Dedicated embedding models become available with their own vignette, and several models can be queried at once — an early sign of the comparative, batch-oriented workflow the package later built out.

    View source ↗