Thoriumi/Fuuga-Laya-emotion-laya

Laya, retrained with Fuuga on Emotion (Laya format)

What this is. Laya, a typed-decision model (ModernBERT-large encoder, 395M parameters, with its own two-layer decision head), imported whole into Fuuga with fuuga import-laya, fine-tuned, head and encoder together, on DAIR Emotion (6 classes), and written back with fuuga export laya in Laya's own layout: the same files Laya's repository has, for Laya's own code to load.

Why it exists. A test case for Fuuga's Laya import and export pair: import a published decision model exactly, retrain it on labelled data, measure it on the same rows before and after, and hand it back in the format it came in. The same model with its encoder as Transformers' ModernBertModel is the companion repository Fuuga-Laya-emotion.

Files

File
model.safetensors	the encoder under encoder., the decision head under Laya's names (type_emb, head.layers.*, scorer.*, act_head.*) and the temperature buffer, F32 (1.7 GB)
encoder/config.json	the ModernBERT config Laya builds its encoder from; norm_bias is true, because fine-tuning moved the LayerNorm biases, which are written
tokenizer/	Laya's tokenizer, unchanged
rl_agent_config.json	Laya's own settings (encoder name, act costs, precision) with this model's head size, input budget and temperatures; fuuga_act_head_current: false (see below)
FUUGA_LAYA_EXPORT_NOTES.md	what the export wrote and from where
eval/	the evaluation reports of every model below, the paired comparisons and the training logs

Training

Data	DAIR Emotion (dair-ai/emotion): a fixed-seed sample of 4,000 training rows and 2,000 test rows from the published splits; the question "Which emotion does this message express?" over sadness, joy, love, anger, fear, surprise
Start	fuuga import-laya of Laya's root checkpoint, checked against Laya's own outputs (identical token ids and option markers, logits within 4e-6)
Method	fuuga decide train --decision-model: 3 epochs, batch 16, head at 1e-4, encoder at 2e-5, loss the log score, 512-token input budget (192 for question and options)
Hardware and time	RTX 4070 12 GB; 6.5 minutes

Results

Held-out accuracy on the 2,000 test rows (chance 16.7%):

Model	Accuracy	NLL	Brier	ECE	Acts at confidence ≥ 0.90
Laya zero-shot, with its own temperatures	58.7%	1.963	0.702	0.305	68.3% of rows, 66.2% correct
Laya, fine-tuned (this model)	89.9%	0.525	0.173	0.074	92.3% of rows, 93.6% correct

Paired on the same rows, this model is right on 679 of the 734 rows the two decide differently (McNemar's exact test, p < 0.0001). The eval/ reports also hold Laya with only its head retrained (58.6%) and the stock ModernBERT-large trained the same way (90.0% fine-tuned, a tie with this model).

Using it

With Laya's own code, point it at this folder as at Laya's. Two differences from Laya:

• Temperatures are 1. The fine-tuned model has no fitted temperature table yet; its probabilities are its raw ones (ECE 0.074 on the test rows). Fit one on held-out rows (Fuuga's decide calibrate, then fuuga export laya --calibration) before relying on them.
• The act head is Laya's, not refitted. Fuuga retrains the option scorer but not the act head, so the act probability Laya reports does not describe this scorer; use the option probabilities. rl_agent_config.json records this as fuuga_act_head_current: false, which Laya ignores and Fuuga reads.

With Fuuga, fuuga import-laya --donor <this folder> --output laya-emotion gives the decision model back; fuuga decide predict --decision-model laya-emotion ... or fuuga-serve --decision-model laya-emotion answers with it.

Verification

• Round trip in Fuuga: this folder re-imported with fuuga import-laya gives the trained model's prediction on all 2,000 test rows, byte for byte (examples/showcase/export-laya.ps1 -Format laya).
• The export is the reverse of the import: Laya itself imported and exported without training gives back all 206 of its tensors value for value and its temperatures, and the re-import passes Fuuga's check against Laya's own outputs.
• Not yet checked: loading this folder in Laya's own Python code. The files follow what Laya's loader reads (weights loaded strictly, the encoder built from encoder/config.json, the act head sized from act_costs), but no run of Laya's code on this folder has been made.

Limitations and terms

• Fine-tuned on this one question only: Laya's other decision tasks were not measured after training and are likely degraded.
• The weights derive from Laya (convaiinnovations/laya, Apache 2.0) and are released under the same licence, with the changes described above.
• The training data, DAIR Emotion, is for "educational and research purposes only" (its dataset card), so this model is released for research and evaluation. If you use it, cite the dataset: Saravia et al., "CARER: Contextualized Affect Representations for Emotion Recognition", EMNLP 2018.