arabellako22/chyon_merged_full_gemma_4_legacy
The arabellako22/chyon_merged_full_gemma_4_legacy is a 5.1 billion parameter conversational text generation model based on the Gemma 4 architecture, developed by arabellako22. This model, with a 32768-token context length, was created by merging a LoRA adapter with the unsloth/gemma-4-E2B-it base model. It is primarily intended for Korean language tasks, though it is not recommended for use due to a recorded issue of repeated prompt-template tokens in generated output.
Loading preview...
Overview
The arabellako22/chyon_merged_full_gemma_4_legacy is a 5.1 billion parameter model built on the Gemma 4 architecture, developed by arabellako22. It was created by merging a LoRA adapter (arabellako22/chyon_lora_model) with the unsloth/gemma-4-E2B-it base model using the standard PEFT merge_and_unload() path. This model is designed for conversational text generation, primarily in Korean, and has a context length of 32768 tokens.
Key Limitations and Status
This model is not recommended for use. Artifact testing revealed a significant issue: repeated prompt-template tokens in generated output during persona prompt generation. Consequently, this version failed its post-merge smoke test and was rejected in favor of a verified Unsloth re-merge. The repository is retained solely for reproducibility and comparison purposes.
Good for
- Reproducibility and comparison: Useful for researchers or developers who need to examine the original merge that exhibited issues.
- Understanding merge failures: Provides a concrete example of a model merge that did not pass quality checks.
Not recommended for
- Inference or deployment: Due to the identified issue of repeated prompt-template tokens.
- GGUF conversion: Should not be used as a source for GGUF builds.
- Production environments: Use the verified replacement model instead.