Log
Training rounds, results and lab news, newest first. Entries are short and factual; the detail lives on the page each one links to.
Ymir Labs has a home: what we’re researching, what Candid is meant to do, where we are, and REN’s results in full.
This site sets out the question the lab works on, the outcomes Candid is designed for, and an honest account of our stage: early, pre-product and pre-revenue.
Every claim on it carries a label that says whether it has been measured, is in research, or is planned. We will keep this log as the dated record of what changes.
The sixth training round of REN ran on the same free hardware as the first five.
Round six was trained on Kaggle’s free tier (2× NVIDIA T4), like every round before it. Its runs and evaluations are on the public training log.
The figures quoted on this site are from round five.
Evaluated on a held-out test set against the untouched base model, with 95% bootstrap confidence intervals.
Against the untouched Qwen3.5-9B base model, evaluated in the same session, REN was clearly better on terminal tasks (42/42 against 36/42), quick judgements, calibration, structured-format answers, format compliance and tool-using agent turns. General knowledge showed no clear change, which means no loss.
Handling the interruption itself moved from 33/42 to 35/42. The confidence interval includes zero, so we count it as within noise: not a win yet.
The round ran for 740 optimiser steps over 20,441 examples and 9.51M tokens. The session cap cut it at step 646; it resumed from checkpoint with no skipped or NaN steps.
REN is a solo proof of concept. Candid is a separate architecture, to be built from scratch once the lab is funded.
The first four rounds of REN, a solo proof of concept trained and evaluated on free GPUs.
REN is a LoRA fine-tune of the open Qwen3.5-9B model, trained so it can be interrupted mid-answer and keep working sensibly. Its first four rounds ran on Kaggle’s free tier.
REN is a solo proof of concept by the founder. It is not Candid, and Candid is not built on it.
Every run and evaluation is logged publicly on the training log.
