ssadqdsacf/cross-unlearning-case6-qwen35-4b-retain-oracle-1epoch-epoch10-lineage
The ssadqdsacf/cross-unlearning-case6-qwen35-4b-retain-oracle-1epoch-epoch10-lineage model is a 4.5 billion parameter language model based on the Qwen3.5 architecture. It functions as an MLLMU-Bench-style retain-only Oracle, specifically trained for one epoch from an original base model. This model is designed for specific evaluation scenarios, particularly for classification, mask, and generation tasks with fixed 4-shot greedy decoding.
Loading preview...
Model Overview
This model, ssadqdsacf/cross-unlearning-case6-qwen35-4b-retain-oracle-1epoch-epoch10-lineage, is a 4.5 billion parameter variant of the Qwen3.5 architecture. It is configured as an MLLMU-Bench-style retain-only Oracle, having undergone a single epoch of training from its original base model. The repository provides both a directly loadable merged model and its corresponding LoRA adapter.
Key Characteristics
- Architecture: Based on the Qwen3.5-4B profile, specifically
qwen3.5-4b-profile-400-sft-v2-r32-cosine-15e-selected-epoch10. - Training: Trained for exactly one epoch, with SFT selection from complete epoch 10 of a longer training run.
- Evaluation: Utilizes fixed 4-shot greedy decoding. Classification tasks employ native thinking, while mask and generation tasks disable thinking.
- Purpose: Primarily intended for specific evaluation contexts, particularly for three-direction classification, mask, and generation summaries.
Intended Use Cases
- Research & Evaluation: Suitable for researchers studying model unlearning, retention, and specific MLLMU-Bench style evaluations.
- Comparative Analysis: Can be used as a baseline or comparison point in experiments involving model lineage and training protocols.
This Oracle model is evaluated separately and is not part of the formal six-model matrix mentioned in its provenance.