QELM Research Notes

QELM 0.1.4 is a public hybrid language-model framework with trainable quantum channels, amplitude-and-phase features, attention and feed-forward paths, next-token training, saved-model tools, and QSA 0.2 exact execution.

Research brief

Prepare token state, execute trainable hybrid model blocks, produce next-token logits, train with parameter-shift or SPSA and Adam, and compare CPU, Aer, IBM-compatible, or certified QSA routes.

Problem: Hybrid language-model experiments need trainable model paths, repeatable execution, explicit backend boundaries, and evidence that can be traced to a named workload.
Hypothesis: Compact trainable quantum channels can contribute useful language-model computation when compared under controlled datasets, baselines, and execution settings.
Current boundary: QSA's fast exact routes depend on certified structure, arbitrary dense states remain exponential, and no published result establishes universal quantum advantage or production-LLM performance.

Quantum language modeling

Study Overview

QELM's public framework can prepare datasets, create and train compact models, save and load model state, generate next-token logits, and run local trainer and saved-model chat tools.

Trainable quantum channels and sub-bit amplitude-and-phase features participate in attention and feed-forward paths rather than sitting outside the language-model loop.

QSA 0.2.0 is the compiled exact execution layer beneath supported QELM workloads; Brain supplies protected orchestration, memory, action, learning, and evaluation around the model.

Public package: QELM 0.1.4 for Python 3.9 and newer.
Native layer: QSA 0.2.0 with C++20, C ABI 1.5, Python bindings, and QSC v1.
Protected layer: current Brain model state and private knowledge are not distributed.

Variables

Language input

Dataset and token state

Dataset split, token maps, context length, and embeddings define what the model can learn and what a comparison means.

Evaluation Lens

Choose which part of a QELM run to inspect.

Language behavior

Inspect the dataset, token map, context, model blocks, optimizer, seed, and held-out outputs.

Study record

Build

Public model framework

QELM 0.1.4 creates, trains, saves, loads, and runs compact hybrid language models.

Execute

Exact native runtime

QSA 0.2.0 supplies certified structural execution, reusable plans, estimators, sweeps, and gradients.

Integrate

Protected Brain system

Brain adds governed memory and action, learning, recovery, orchestration, and controlled evaluation around QELM.

Next experiments

01
Complete full sequential execution across configured QELM blocks.
02
Evaluate independent amplitude and phase learning paths.
03
Broaden QSA 0.2 estimator, gradient, and model-execution integration.
04
Publish larger controlled language runs with held-out data and matched baselines.

Related research area All studies