Move text-gen files to finetune/text/; drop finetune tests

Merge branch 'vivamais-vision-v1' into vision-into-main

Consolidate vision finetune files under finetune/vision/

Force train GPU via decorator env, not with_options

Add Modal dataset publisher + refresh dataset card

Add v5 train config (restraint data, 2 epochs, no in-train eval)

Generate the synthetic dataset on Modal, not the local box

Add Stage-2 field diagnostic: split recall vs over-fill

Make sharded eval robust to HEIC and per-shard failures

Reject weak Rio v4 travel teacher rows

Clear deterministic generation sampling flags

Pin Rio v4 as standalone Modal module

Use A100 for Rio v4 text run

Add Rio v4 text safety recipe

Add v3 text dataset publisher

Parametrize merge/GGUF repo names + v4 export config

Shard the vision eval across GPU containers + full real holdout

Add synthetic-only v4 train config + A100/skip-upload wiring

Pass HF token to eval container for private merged repos

Log per-row progress in the vision eval pipeline

Upload merged v3 weights from the Modal container, not the local box

Fix v3 GGUF publish: drop bad smoke flag, pass HF token to container

Reject Rio v3 teacher meta answers

Inline workspace guard in train/merge so remote containers import

Diversify Rio v3 teacher prompts

Tighten Rio v3 teacher recipe

Add Rio-distilled Viva Mais v3 pipeline

Render PDFs, make teacher loop crash-proof, fewer shards

Raise teacher Stage-2 token cap for multi-segment docs

Add Rio 3.1 teacher eval harness

Force teacher Stage-2 JSON via assistant prefill

Handle Rio as a reasoning model in the teacher pipeline

Parallelize and instrument the teacher pipeline

Wire v3 train, merge, GGUF export, and runtime selection

Add accent-sensitive vision eval harness

Add Modal teacher and judge pipelines (Rio-2.5-Open-VL)

Purge legacy v1 finetune/eval; re-ground on current domain

Add Rio vLLM teacher serving bridge

Add Rio model feasibility inspection

Add Rio teacher endpoint probe

Add shared PT-BR conversation eval suite

Expand Viva Mais text eval suite

Add broad MiniCPM5 text v2 distill pipeline

Handle Unsloth GGUF output directory

Add MiniCPM5 text GGUF export

Add Viva Mais travel vocabulary SFT bucket

Add Modal PT-BR QA dataset inspector

Align MiniCPM5 v1 QA grounding behavior

Allow MiniCPM5 evals to use thinking mode

Make MiniCPM5 SFT mix weights configurable