peerproblem/Kanana-IRI-3B-QLoRA

Kanana IRI 3B QLoRA version history

Powered by Kanana. This repository preserves all five LoRA adapters produced while improving the IRI Korean child-conversation prototype. The root adapter is the selected v5 release. Every historical adapter remains under versions/ so the full iteration history is auditable.

Recorded v5 results

Check	Recorded result	Scope
Training and validation data	412 training / 92 validation examples	Three-epoch training run
Best validation loss	2.103572	Lowest recorded value among v2 through v5; validation sets differ
Fresh-process generation	5/5 selected examples generated	Adapter reload and generation check, not an accuracy or safety pass rate
Targeted correction checks	Corrective responses observed for pet-harm and disability-exclusion requests	Two selected examples from the five-example check
Independent final evaluation	Pending	No v5 holdout score is available

The selected checkpoint was restored and reloaded in a fresh process. These results document completed training and targeted generation checks. See the project assessment for the selection evidence and evaluation scope.

Version sequence

Version	Train rows	Validation rows	Best validation loss	Outcome
v1	50	12	n/a	historical baseline adapter
v2	268	62	2.2475779056549072	retrained candidate
v3	332	76	2.1451010704040527	full-regression candidate rejected
v4	372	84	2.1469812393188477	supplemental-holdout candidate rejected
v5	412	92	2.103572130203247	selected; fresh-process generation completed for 5/5 examples

v2 through v5 used the same three-epoch schedule while adding reviewed correction data. Their validation sets differ, so the recorded losses do not establish a like-for-like response-quality improvement. Behavioral selection used separate scenario evaluations. A lower validation loss alone did not determine the release.

v3 was rejected after the 300-scenario regression exposed three severe raw-response safety failures. v4 was rejected after a fresh 60-scenario holdout found unsafe content in exclusion and pet-harm request families, with 46/60 guarded action matches. v5 achieved the lowest recorded validation loss and passed fresh-process loading plus generation checks on those failure families.

A separate v5 holdout run did not start because three Runpod allocation paths failed to provide a GPU host. This limitation is part of the release record. v5 is the best available candidate, not a fully validated child-safety model.

Loading the selected adapter

Use kakaocorp/kanana-2-3b-instruct at commit 6a5d7889964c4c590299d16e309eabab1f73f8a9 and load this repository with PEFT. To reproduce a historical version, download the matching versions/vN directory and pass that directory to PeftModel.from_pretrained.

The adapters do not include the base weights. Input checks, output checks, deterministic safety routing, rate limits and service policy live in the IRI application and are not embedded in the adapter. Do not expose a raw adapter endpoint as a child safety product.

Reproducibility and license

Each version directory contains its adapter files, a machine-readable metadata record and the available training manifest. SHA-256 values are listed in version-index.json. The base Kanana Open License Agreement in LICENSE applies to these derivative adapters. NOTICE contains the required attribution. Kakao did not endorse this project.