Qualia is your autonomous AI researcher.

Use it to autonomously run experiments, interpret results and propose new directions, on your machine or in the cloud.

lm-training

(10)

Tasks

Knowledge

Baseline establishes the comparison point

The 124.4M-parameter transformer reaches a validation loss of 2.824 on WikiText-103. All three variants use the same validation set and training budget.

Sources

FigureBaseline training curve

Training loss decreases over the fixed training budget.

Rotary embeddings improve validation loss

RoPE with a 2,048-token context lowers validation loss to 2.741, a 3.0% improvement over the baseline. This experiment changes both position encoding and context length, so their individual effects remain untested.

Sources

Warmup and shared embeddings beat the baseline

Adding 1,000 warmup steps and tying input and output embeddings lowers validation loss to 2.709, a 4.1% improvement under the same training budget.

Sources

SwiGLU and RMSNorm deliver the largest gain

Gated feed-forward layers with RMS normalization achieve a validation loss of 2.641, down 6.5% from the baseline and the lowest of the three tested variants.

Sources

SwiGLU and RMSNorm lead this comparison

All three variants outperform the baseline. SwiGLU with RMSNorm is the strongest tested combination; combining it with warmup or rotary embeddings remains a follow-up experiment.

Sources

TableTransformer experiment leaderboard

Variant	Val. loss	Gain
Baseline	2.824	—
RoPE + longer context	2.741	3.0%
LR warmup + weight tying	2.709	4.1%
SwiGLU + RMSNorm	2.641	6.5%

FigureValidation loss comparison

SwiGLU + RMSNorm finishes with the lowest validation loss.

SwiGLU + RoPE

Validation loss: 2.583

Sources

SwiGLU + warmup

Validation loss: 2.562

Sources

Combine all improvements

Validation loss: 2.498

Sources

Learning-rate sweep

Validation loss: 2.607

Sources

Regularization sweep

Validation loss: 2.619

Sources

Repeat across three seeds

Validation loss: 2.646 ± 0.008 across three seeds

Sources

Compare follow-up experiments

Combined model: 2.498 · 11.5% below baseline

Sources

Train 125M transformer

Find three ways to improve this transformer and test them in parallel.

The baseline is trained. I’ll compare three improvements on the same validation set:

RoPE + longer context
LR warmup + weight tying
SwiGLU + RMSNorm

Each agent will train its variant and record the results in the task map.

Started 3 agents

All three agents responded. Let's aggregate the data and report results.

Made 1 change

SwiGLU + RMSNorm has the lowest validation loss: 2.641, down 6.5% from the baseline.

RoPE reached 2.741 and warmup with weight tying reached 2.709. The comparison plot and all three experiments are saved in the notebooks.

Go one level deeper. Combine the improvements, tune the winner, and check that the results hold up.

I’ll launch six follow-up agents: combine SwiGLU with RoPE, add warmup, combine all improvements, sweep learning rates, tune regularization, and repeat the winner across three seeds.

Started 6 follow-up agents

All six follow-up agents finished. I’m comparing their results with the first-round winner and checking the seed runs.

Combining all improvements wins: validation loss 2.498, down 11.5% from the baseline.

SwiGLU + RoPE: 2.583
SwiGLU + warmup: 2.562
All improvements: 2.498
Learning-rate sweep: 2.607
Regularization sweep: 2.619
Original SwiGLU winner, three seeds: 2.646 ± 0.008

The seed runs support the first-round gain. The combined model is the next candidate to validate across seeds; all six experiment notebooks are saved.

Qualia supports every step of the research process.

Process and work with data

Unify diverse datasources across files, datasets, and the web
Autonomously clean and preprocess data

FileEditSelectionViewGoRunHelp

TAM Research

TasksKnowledgeKernels

FILES

datanow

tam_companies.csv218 MB

industry_segments.parquet

analysisnow

01_load_tam.ipynb

02_profile.ipynb

03_normalize.ipynb

04_deduplicate.ipynb

08_benchmark.ipynb

AGENTS (0)(1)(2)(3)(4)(5)(6)(7)(8)

Load TAM datasetnow

Reading 2.4M rows01_load_tam.ipynb

0/1 tasks completed1/1 tasks completed

Profile and validatenow

Checking 38 columns02_profile.ipynb

0/1 tasks completed1/1 tasks completed

Normalize company namesnow

Standardizing entities03_normalize.ipynb

0/1 tasks completed1/1 tasks completed

Resolve duplicatesnow

Matching company records04_deduplicate.ipynb

0/1 tasks completed1/1 tasks completed

Enrich market segmentsnow

Joining industry data05_segments.ipynb

0/1 tasks completed1/1 tasks completed

Optimize the pipelinenow

Benchmarking lazy queries06_optimize.ipynb

0/1 tasks completed1/1 tasks completed

Validate outputnow

Running quality checks07_validate.ipynb

0/1 tasks completed1/1 tasks completed

Compare performancenow

Comparing pipeline runs08_benchmark.ipynb

0/1 tasks completed1/1 tasks completed

01_load_tam.ipynb08_benchmark.ipynb

02_profile.ipynb

04_deduplicate…

Python 3

TAM dataset

Load, clean, and validate the total addressable market data.

[2]

preview = tam.select(
    "company_name", "country", "industry", "revenue_usd"
).head(5).collect()
preview

5 rows × 4 columns

	company_name	country	industry	revenue_usd
0	Northstar	US	Software	42,500,000
1	Meridian	GB	Healthcare	18,200,000
2	Clearwater	US	Energy	67,800,000
3	Evergreen	DE	Industrials	31,400,000
4	Atlas	CA	Software	24,600,000

[8]

plot_pipeline_comparison(baseline, optimized)
plt.show()

3.1× faster. Same validated data.

Lazy execution, parallel joins, and a reusable Parquet pipeline.

Code＋ Markdown

Load TAM dataset

load the TAM dataset

Thought (2.37s)

I’ll load the TAM dataset and coordinate agents to profile, clean, and validate it in parallel.

Read 2 files and created 7 agents

To AgentProfile and validate

Profile missing values and validate the schema across all 38 columns.

To AgentNormalize company names

Normalize company names, country codes, and revenue types.

To AgentResolve duplicates

Resolve duplicate company records and preserve the latest observations.

To AgentEnrich market segments

Join the industry segments and report coverage of the cleaned data.

To AgentOptimize the pipeline

Optimize the pipeline with lazy execution and parallel joins.

To AgentValidate output

Check unique IDs, valid revenue, and output schema before saving Parquet.

To AgentCompare performance

Compare three benchmark runs and confirm identical validated output.

Edited 12 cells, ran 8 commands

The data is cleaned and validated. I’m checking a faster pipeline with lazy execution and parallel joins.

Compared 3 benchmark runs

The optimized pipeline is 3.1× faster, with identical validated output.

Runtime: 192s → 62s
Peak memory: 4.8 GB → 1.3 GB
Quality checks: all passed

Notebook: 08_benchmark.ipynb

0/1 tasks completed1/1 tasks completed

Explore, build, validate, / for skillsExplore, build, validate, / for skills

Illustrative workflow: load the TAM dataset, coordinate eight agents, prepare data in Python notebooks, and compare a 62-second pipeline against a 192-second baseline. Dataset and benchmark values are simulated.

Build recurring pipelines

Run workflows on a recurring schedule to generate live reports
Autonomously refine pipelines based on feedback to improve future runs

Build and train ML models

Autonomously train models and improve existing ones, from linear regression to deep neural networks
Easily post-train 10B-parameter foundation models on your data for specialized tasks

Explore and test new ideas

Fan out agents to test features, architectures, and training strategies in parallel
Diagnose model failures and autonomously propose and run follow-up experiments

FileEditSelectionViewGoRunHelp

Housing Research

TasksKnowledgeKernels

FILES

experiments

housing.csv

baseline.ipynb

comparison.ipynb

error_analysis.ipynb

AGENTS

Feature engineeringnow

features.ipynb

Test new feature combinationsTask completed

Gradient boostingnow

boosting.ipynb

Compare learning ratesTask completed

Random forestnow

forest.ipynb

Tune depth and leaf sizeTask completed

Neural networksnow

neural.ipynb

Explore network architecturesTask completed

Error analysisnow

errors.ipynb

Inspect model failure casesTask completed

Feature ablationnow

ablation.ipynb

Isolate useful featuresTask completed

Regularizationnow

regularization.ipynb

Test regularization strategiesTask completed

Ensemblingnow

ensemble.ipynb

Combine independent modelsTask completed

Calibrationnow

calibration.ipynb

Check prediction calibrationTask completed

Data qualitynow

quality.ipynb

Audit missing data and outliersTask completed

Cross-validationnow

validation.ipynb

Check stability across splitsTask completed

Spatial featuresnow

spatial.ipynb

Test geographic contextTask completed

baseline.ipynb

Tasks

Tasks

MapLensTimeline

Show knowledgeColor by metricNew task

Experiment coordinator

Establish the baseline

T-demo00000000

Prepare the dataset and establish a reproducible baseline for all experiments.

FindingsFixed validation split recorded.▸ Figure: Baseline predictions

Data quality

In progress

Audit missing data and outliers

T-demo00000010

Run the experiment on the agreed split and capture the comparison in quality.ipynb.