Fenua model training

Stage B bilingual adapter

A clearer view of the model run: loss, generation evals, dataset mix, and the community signals feeding the next Tuvaluan-English iteration.

TVL Stage B local snapshotFriezaForce/tvl-en-llm-translation-stage-b
Showing a packaged local snapshot because the live training backend is not reachable in this environment.
Train NLL
1.3800down 7.4%

Step 8,400

Validation NLL
1.6300

Step 8,000

chrF++
41.8

Step 8,000

BLEU
18.4

26.0% exact

Last updatedMay 15, 12:00 AM
CheckpointBest eval snapshot at 8,000
Sampler step8400

Loss curve

Training is moving in the right direction

Train Validation
1.302.142.993.8402,1004,2006,3008,400

Dataset

What the model is learning from

342,000 examples
synthetic tvl128K
crosslingual84K
english62K
anchor42K
real tvl chat26K

Evaluations

Generation quality

StepchrF++BLEUExact
8,00041.818.426.0%
6,00037.815.621.0%
4,00031.210.914.0%
2,00024.66.88.0%

Coverage

Task families

134Ktranslation
72Kchat
54Kqa
36Ksummarization
26Kgrammar
20Ksafety