Skip to content

Releases

Franok v0.4.0

Choose a tag to compare

@github-actions github-actions released this 19 Nov 07:32

Franok 0.4.0

Full-parameter fine-tune of 0.3 on 35,672 teacher-written, programmatic and earlier conversations,
plus passage reading. Results: docs/evaluation/v4/final.md

Franok v0.3.0

Choose a tag to compare

@github-actions github-actions released this 18 Nov 17:00

Franok 0.3.0

This version continues the project's own 314M-parameter English–Turkish model on 383,385,600 new text targets, then applies supervised training, two LoRAs and on-policy reinforcement learning. The Ubuntu package runs Q8 locally and opens from the Ubuntu dock. On first use it downloads 353.4 MB of weights. Saved sessions work offline after that download.

A bilingual session search can reopen and cite an old turn. An optional passage reader can quote the exact source text. The Architecture view lets you inspect a real token's route and its predicted next pieces.

General conversation remains experimental. Franok passed 0/32 held-out conversation turns. See scores and transcripts, browser chat and model weights.

Franok 0.2.0 — English and Turkish local chat

Choose a tag to compare

@yunusemrejr yunusemrejr released this 18 Nov 11:46

Franok is an experimental bilingual language model trained from random initialization. Each token activates roughly 50 million of its 314 million parameters.

Download the amd64 .deb below for Ubuntu. The first conversation downloads 353 MB of Q8 weights. Later conversations work offline, and saved sessions stay on the device. The browser uses a 196 MB Q4 model.

This release extends the neural context to 1,024 tokens and includes full supervised tuning, two supervised LoRAs and an on-policy reinforcement-learning LoRA. The conversation helper handles greetings and clarifications. Direct multi-turn generation remains unreliable. Q4 and Q8 agreed with FP32 on 122 and 128 of 128 held-out answer choices, respectively.

Browser chat · Weights and adapters · Architecture · Evaluation and transcripts · Watch the two-minute film