Karti-Small-RSI-3B

Karti-Small-RSI-3B is Lumbridge's compact local model family for Veronica and Prime Agent workflows, with reliable tool calling and offline operation. It begins with the fully open SmolLM3-3B foundation and improves through a recurring loop of reviewed traces, verifiable rewards, bounded training, held-out evaluation, and explicit human promotion.

The first private checkpoint, Karti-Small-RSI-3B-2026.08-W35, completed its initial LoRA SFT run on August 26, 2026: 64 optimizer steps over 512 verified rows. That first cycle established the complete path from curated data to a trained checkpoint and clean evaluation. Each result now becomes evidence for the next data, reward, and training pass.

This public repository is the reference page for the model program: its open foundation, adaptation recipe, improvement loop, and progress over time. Karti deployment artifacts and the private data that teach them remain under Lumbridge control.

Program status

Program element	Status
Karti model family	Active
First private LoRA checkpoint	Training complete
Reviewed-data and evaluation loop	Operational
Public model card and recipe	Published
Next improvement cycle	In preparation
Release direction	Monthly checkpoints

Recursive self-improvement

RSI is a practical, measured loop:

1. Collect reviewed public seeds and authorized Veronica or Prime Agent traces.
2. Verify rights, redaction, provenance, response structure, and tool intent.
3. Train a bounded candidate from a pinned open foundation.
4. Evaluate held-out tool behavior, clarification, response contracts, and confirmation boundaries.
5. Improve the next mixture and reward design from the measured evidence.

The loop is recursive because every checkpoint helps identify the highest-value examples for the checkpoint after it. It stays human-directed and promotion-gated, so capability grows together with trace quality, evaluation coverage, and operational confidence.

What we are building

Karti is designed to run locally beside ASR and TTS as part of Veronica and Prime Agent workflows, including during an internet outage. The family is being shaped to:

• ask concise clarifying questions when important information is missing;
• emit exact, schema-valid tool calls;
• preserve confirmation and proposal-only boundaries;
• work naturally with Lumbridge's local agent tools;
• respond quickly enough to feel at home in a voice-assistant loop; and
• improve from reviewed real-world traces without publishing household context.

The first evaluation cycle highlighted clarification coverage and strict response-contract structure as high-value areas for the next pass. That is the loop working: measured behavior becomes targeted data, and targeted data becomes a stronger generation.

Open foundation

• Base model: HuggingFaceTB/SmolLM3-3B
• Pinned base revision: a07cc9a04f16550a088caea529712d1d335b0ac1
• Base license: Apache-2.0
• Adaptation: BF16 LoRA supervised fine-tuning with TRL
• First-cycle context: 2,048 tokens
• Tool representation: Structured XML/JSON tool calls
• Initial objective: precise, proposal-only tool behavior

SmolLM3 supplies the open base weights, training code, intermediate checkpoints, and upstream training documentation. Karti builds a recurring local-agent improvement program on top of that foundation.

Adaptation sequence

1. Select capped, revision-pinned utterance seeds from reviewed public sources.
2. Privately re-author and verify proposal-only demonstrations.
3. Add authorized, redacted, deduplicated private and reviewed synthetic traces with recorded provenance and rights.
4. Freeze the holdout before training.
5. Train a bounded LoRA candidate from the immutable base revision.
6. Evaluate tool names, order, arguments, response contracts, and confirmation boundaries through Prime Intellect Verifiers and Lumbridge Tera.
7. Promote after held-out improvement and owner review.

The machine-readable public portion of the process lives in recipe.json.

Data foundations

The public bootstrap can draw reviewed utterance seeds from:

• acon96/Home-Assistant-Requests
• NousResearch/hermes-function-calling-v1
• Team-ACE/ToolACE

Karti re-authors and verifies its training targets. Redaction, rights checks, deduplication, provenance, and holdout isolation happen before a row becomes training eligible.

Safe tool behavior

The initial home-assistant policy exposes bounded read/list tools and a proposal tool. Confirmation remains outside the model's callable path. This makes safe behavior part of the reward contract and part of every promotion decision, not an instruction added after training.

Follow the program

The model family, monthly loop, and future Karti generations live at lumbridgecorp.com/models.