Laya Instinct
Laya Instinct is a fine-tuned Laya decision model for choice, score, and noul (yes/no) questions. It is a full checkpoint, not an adapter. The weights are released under Apache-2.0, the same license declared by the base model.
Training data and attribution
The 500,000-record training mixture was constructed from the following Hugging Face datasets. Counts are records in the local training split, not the total size of each upstream repository.
| Source dataset | Training records | Contribution |
|---|---|---|
| tasksource/tasksource-jev-typed-decisions | 221,689 | Direct normalized records |
| SargeDev/jev-distill-corpus-v3 | 193,978 | Human/teacher decision targets; exact-uniform yuri_v1 placeholders excluded |
| Praveenrajus/jev-bench | 55,422 | Normalized benchmark records |
| LocalLLaMA/typed-decisions | 1,200 | Typed decision workflows |
| Tasksource-derived robustness augmentation | 27,711 | Option-order perturbations of Tasksource records, not an additional upstream dataset |
The mixture contains 504,800 individual decision questions. A separate 6,000-record calibration split was used to fit per-decision-type temperatures. The validation split was excluded from optimization and calibration. The dataset builder and its local manifest.json record the construction settings; upstream test/OOD splits were excluded.
Data licensing: Apache-2.0 here describes the released model weights; it does not relicense the training data. The Tasksource collection has source-specific licenses and usage terms, including commercial/non-commercial/unspecified labels. jev-bench aggregates datasets with their own licenses (for example, ARC rows identify CC-BY-SA-4.0). Review the linked dataset cards and original source terms before using this model or reconstructing the mixture for a particular purpose. The other two linked datasets declare Apache-2.0 on Hugging Face.
Training and evaluation
Starting from convaiinnovations/laya, the model was trained for four epochs on one RTX 3090 Ti, with effective batch size 64 and maximum sequence/head lengths of 1024/256. The final checkpoint uses calibrated temperatures for choice, score, and noul decisions, respectively: 4.1300, 2.5513, and 3.4845.
On a 6,000-record held-out validation split from the same source families, evaluated with a common 1024/256 inference budget:
| Model | Accuracy | Brier score | Log loss |
|---|---|---|---|
| Plain Laya | 43.48% | 0.5714 | 1.7356 |
| Laya typed-decisions | 43.78% | 0.5010 | 1.3870 |
| Laya Instinct | 71.25% | 0.3442 | 1.0567 |
This is an in-domain validation result, not an independent downstream benchmark. High-cardinality classification remains weak (Banking77: 5%; CLINC150: 1%). The two official checkpoint comparators emitted a temperature-bucket warning, so their calibration-error values are not directly comparable; accuracy and the metrics above are reported under the same inference budget.
Use
Install the Laya package and load this checkpoint with laya.Agent("chand1012/laya-instinct", device="cuda"). The repository contains model.safetensors, rl_agent_config.json, encoder configuration, and tokenizer files. It is not a standard generative language model.
Reproducibility
- Model file SHA-256:
584c72a0f320c2884deacba31fb4d4378f548f264549bde841086a02857248b6 - Mixture manifest SHA-256:
f1f89efa5d6572c47a8bae752ed2f93738f83c962c9f258b1d09f4bcab1530b7 - Training seed:
20260922; dataset construction seed:20260926