REN · Proof of concept

What one person can do on free GPUs.

REN shows that one person with zero budget can train, evaluate and publish a model rigorously. It has been trained in six rounds so far, all on Kaggle’s free tier (2× NVIDIA T4), and every run and evaluation is logged in public.

MeasuredPublic training log

REN is not Candid.

REN is a solo proof of concept by the founder. It is not Candid, not a prototype of Candid, and Candid is not built on it. Candid is a separate architecture, to be built from scratch once the lab is funded. REN is here because it shows how we work.

01What it is

A model you can interrupt.

REN is a LoRA fine-tune of the open Qwen3.5-9B model. It is trained so that it can be interrupted mid-answer, told to stop, corrected, or given new information, and keep working sensibly.

The model weights are private. The training runs and evaluations are not: they are all on the public training log.

02How we tested it

Same session, same test, honest intervals.

The figures below are from round five. Each comparison follows the same rules.

Test set: Held out from training, and checked for overlap with the training data.
Baseline: The untouched base model, Qwen3.5-9B, evaluated in the same session.
Uncertainty: 95% bootstrap confidence intervals on every comparison.
The bar: A result counts as better only when the interval excludes zero.

03Results

Clearly better on six of eight. Not yet on the one that matters most.

Round five against the untouched base model. General knowledge held steady, which means no loss. Handling the interruption, the thing REN is trained for, improved but sits within noise.

Base Qwen3.5-9B
REN, round five
Scale 0 to 1

REN round five against base Qwen3.5-9B, on a held-out test set, with 95% bootstrap confidence intervals. Scores are proportions from 0 to 1, except calibration, where a lower Brier score is better.
Measure	Base	REN	Verdict
Terminal tasks completedEnd state checked · n=42	0.85736/42	1.00042/42	Clearly better
Quick judgementsAccuracy · n=300	0.677	0.830	Clearly better
CalibrationBrier score, lower is better · n=300	0.419	0.252	Clearly better
Right answer in structured-format tasksn=350	0.654	0.937	Clearly better
Format compliancen=350	0.894	1.000	Clearly better
Tool-using agent turns answered correctlyn=50	0.740	1.000	Clearly better
General knowledgeMMLU-Pro · n=200	0.615	0.630	No clear change (no loss)
Handles the interruptionn=42	0.78633/42	0.83335/42	Within noise: not a win yet

The row we’d most like to be better.

REN handled the interruption in 35 of 42 test cases, against 33 of 42 for the base model. That is a gain of two cases on 42, and the confidence interval includes zero. By our own rule, that is within noise: not a win yet.

We could have left this row off. We didn’t, because it is the measure REN exists to move, and because a results table with only good rows tells you very little.

04Training

Round five, in numbers.

Optimiser steps: 740
Training examples: 20,441
Tokens: 9.51M
LoRA rank: 16
Throughput: ~220; tokens per second
Step time: ~12 h; across two free Kaggle sessions
Skipped or NaN steps: 0
Hardware: 2× T4; NVIDIA, Kaggle free tier

The run, step by step. The free session cap cut training at step 646; it resumed from a checkpoint and finished at step 740, with no skipped or NaN steps.

Session one: steps 1 to 646
Session two, resumed from checkpoint: steps 647 to 740

05Data

Where the data came from.

Openly licensed datasets and programmatic generators.
No outputs from OpenAI or Anthropic models.
Held-out test set checked for overlap with the training data.

Why only one comparison

REN is compared with its own untouched base model, in the same session, on the same test. That is the comparison that isolates what the training did, so it is the only one we report.

Open the training log REN in the lab log

Want to know how this standard carries into Candid? Read the research page.