ARKA Advanced SLM — Direct Full-Weight Multi-Domain Retention v1
This is a standalone model directly fine-tuned from Abhayn01/arka-advanced-slm-base-v2.
No LoRA, PEFT, QLoRA or adapter was used. All model parameters were trainable.
Training domains
- assistant/chat
- general Q&A, commonsense and GK
- science
- defensive security, privacy and safety behavior
- explicitly approved non-sensitive personalization
- previous-model replay/retention
Previous-weight retention
The run first measured the original base-v2 retention loss. During full-weight fine-tuning, FineWeb-Edu, UltraChat and UltraFeedback replay samples were mixed into every training step. When GPU memory allowed, a frozen base-v2 teacher supplied a KL behavior anchor. A fixed retention benchmark was evaluated after every epoch. Only retention-safe epochs were eligible for the final checkpoint; excessive drift triggered early stopping.
Base retention loss: 2.402381
Final retention loss: 2.402381
Retention ratio: 1.000000
Base MMLU label accuracy: 0.2417
Final MMLU label accuracy: 0.2528
Privacy
Raw personal training files are not uploaded. Configured sensitive identifiers and credentials are rejected from personal weight-training inputs.
See artifacts/training_manifest.json, artifacts/dataset_audit.json, and
artifacts/final_metrics.json for details.