KITTYLM // RESEARCH

Independent language-model research

KittyLM

A modular research platform for building, measuring, and understanding compact autoregressive language models from first principles.

Created by KrazyKitty with reproducibility, dataset provenance, secure artifact boundaries, and evidence-led evaluation designed into the foundation.

  • 0.1.0.dev0
  • Python 3.12+
  • PyTorch
  • ROCm research
  • Pre-release
View research status Follow KrazyKitty on Hugging Face
KittyLM abstract modular cat mark

FOUNDATION ACTIVE // NO WEIGHTS RELEASED

01 // Research system

Built to make every result explainable.

KittyLM separates tokenization, architecture, training, inference, evaluation, and governance so each layer can be tested without blurring research claims.

01

Tokenizer

Deterministic byte-level BPE with explicit vocabulary layout, special-token policy, artifact versioning, and lossless byte coverage.

02

Model core

Compact decoder-only Transformer with RMSNorm, RoPE, causal attention, SwiGLU, tied embeddings, KV caching, and explicit parameter accounting.

03

Training

Deterministic configuration, AdamW optimization, precision policy, atomic checkpoints, recovery state, and a fresh-process resume harness.

04

Evaluation

Windowed loss, perplexity, bits-per-byte, controlled generation, throughput measurement, and engineering gates kept distinct from formal experiments.

05

Evidence ledger

Structured records are designed to pin source revision, configurations, dataset identity, seeds, environment, parameter counts, metrics, and limitations.

06

Safety boundary

Secret scanning, artifact controls, integrity-aware checkpoint loading, and explicit authorization gates protect private data and unpublished research.

Research principle: build the smallest trustworthy baseline first.

Architectural experiments only become meaningful when the data, parameter budget, training procedure, evaluation windows, and comparison metrics are controlled.

02 // Evidence snapshot

Engineering foundation, not a model release.

This public snapshot was reviewed against the private default branch on 10 October 2026. Merged code is not presented as trained-model capability.

KittyLM 0.1 milestone status and the evidence each status supports
Milestone Scope Status What can be claimed
Foundation Configuration, security controls, CI design, research ledger Merged Research infrastructure exists in the private source repository.
A Byte-level BPE tokenizer and validation Merged Tokenizer engineering is implemented and tested.
B Transformer baseline, KV cache, parameter accounting Merged Compact model architecture is implemented and tested.
C Training engine, checkpoints, deterministic recovery Merged Training mechanics and recovery infrastructure exist.
D Evaluation, generation, inference measurement Merged Evaluation tooling exists; no model-quality result follows from that alone.
E Local GPU correctness and Nano overfit acceptance Pending Acceptance is not established from the default branch snapshot.
F Curated local-v1 dataset pipeline Planned No approved training dataset exists.
G Formal baseline and tokenizer ablations Planned No formal model-quality results exist.

No KittyLM model weights have been trained or released.

There is currently no KittyLM checkpoint, tokenizer package, dataset, inference endpoint, or downloadable model artifact on this Hugging Face account.

03 // Dataset provenance

Permission and provenance come before ingestion.

No approved KittyLM training dataset exists. Future candidates are not accepted merely because they are public, downloadable, or commonly used.

  1. Identify Record the source, publisher, version, canonical location, and intended research use.
  2. Authorize Review licence terms, permissions, privacy implications, redistribution limits, and project approval.
  3. Validate Check integrity, scan for sensitive material, apply documented filtering, and account for exclusions.
  4. Version Create deterministic splits, bind tokenizer compatibility, and produce a reproducible provenance manifest.

Planned provenance record

  • Source identity, version, licence, and permission evidence
  • Acquisition method, integrity hash, and processing recipe
  • Accepted and rejected document and byte accounting
  • Dataset version, partition assignment, and restrictions

Protected material

  • No credentials, private keys, tokens, or connection strings
  • No private correspondence or confidential project files
  • No automatic training on chats, feedback, or unreviewed interactions
  • No raw, cleaned, or tokenized private corpus in source control

04 // Reproducibility and evaluation

Measurements earn claims.

No formal model-quality results are reported. Engineering checks establish that components behave as intended; they do not establish useful language capability.

Every formal run is expected to pin

  • Source revision and verified configuration hashes
  • Tokenizer and dataset versions with provenance
  • Random seeds, hardware, software, and precision policy
  • Optimizer settings, parameter accounting, and token accounting
  • Checkpoint identity and deterministic resume evidence
  • Metrics, resource measurements, limitations, and failures

Planned baseline evidence

  • Cross-entropy loss and perplexity on controlled splits
  • Bits-per-byte for tokenizer-independent comparison
  • Overfit-gate token-match accuracy
  • Prefill and decode throughput with hardware context
  • Memory use, training stability, and recovery equivalence
  • Known limitations and responsible-use boundaries

Evidence rule: a result that is not in a validated experiment record does not become a research claim.

Hardware-specific smoke measurements remain engineering evidence and are not promoted into benchmark tables or model-card metrics.

05 // Research roadmap

Advance through verified gates.

Roadmap items are research directions, not release dates or capability promises. Each stage depends on evidence from the stage before it.

NOW

0.1 foundation

Close local GPU acceptance, preserve deterministic tooling, and repair any current-head validation drift before experiments.

NEXT

Dataset pipeline

Implement reviewed ingestion, licence and privacy gates, filtering, deterministic splits, and versioned provenance.

THEN

Formal baseline

Train the controlled compact baseline, record reproducibility evidence, and run planned tokenizer comparisons.

LATER

Efficiency studies

Investigate broader data mixtures, stronger evaluation, optimized inference, and carefully controlled architectural alternatives.

R&D

Specialist systems

Explore hybrid sequence mixers, expert routing, reasoning evaluation, retrieval, memory, and secure KittyOS integration.

RULE

No capability by naming

A future version number or architecture proposal never implies that a capability is implemented, measured, or scheduled.

06 // Research ecosystem

One project inside a wider evidence system.

These public companion resources separate dataset governance, agent evaluation, portfolio navigation, and curated study material from KittyLM's model research.

NL

NEURAL LAB

Interactive portfolio map, public Collection explorer, and transparent research-readiness checklists.

 DL

DATA LAB

Documentation-led research for provenance, permissions, privacy, preprocessing, quality, and evaluation resources.

 NA

NEKO//AGENT

Agent architecture, safety boundaries, reproducible trace evaluation, and an offline deterministic demo.

 RC

Research Collections

Four curated tracks covering transformer foundations, model training, data provenance, and agent systems.

Boundary: links describe public research resources, not shared runtime capabilities.

KittyLM still exposes no model weights, inference endpoint, training dataset, or generated output.

Public release gate

What must exist before a model card exists.

KittyLM will publish evidence with artifacts, not ahead of them. Until every release condition is met, this Space remains a research showcase rather than a model repository.

  1. A trained, integrity-checked checkpoint in an approved safe format
  2. A validated experiment record tied to the exact source revision
  3. Dataset and tokenizer versions with complete provenance and licensing
  4. Measured evaluation results, limitations, intended uses, and safety notes
  5. Explicit KrazyKitty approval for public publication