Gamma-30B-A6B-G1-311-HardNegativeScriptMeta-LoRA
This is a G1-308 continuation adapter for Gamma 30B-A6B. It starts from the accepted G1-302 multilingual repair adapter and adds a focused anti-meta, direct-answer behaviour repair.
It is not the final model and not an external expert router. It is a traceable EvoStream adapter chromosome intended to be merged/unloaded or consolidated into a single Gamma candidate after gates pass.
Purpose
Heal the strict multilingual/public-output defect observed after G1-292/G1-305: missing requested script in a few rows and meta-commentary such as "The user wants", "We can say", "I should", or markdown preambles.
Result
- Seed adapter:
GammaAGI/Gamma-30B-A6B-G1-308-AntiMetaDirectAnswer-LoRA - Training status:
OK - Loss before:
1.429807 - Loss after:
1.039675 - Trainable parameters:
110484224 - Trainable ratio:
0.00348658 - Train JSONL path:
/workspace/gamma_30b_alpha_lora/input/gamma_g1_310_hard_negative_script_meta_train_20260623.jsonl - Train items:
55 - Steps:
96 - LR:
8e-06
Next
Run a small targeted smoke gate first, then the balanced strict multilingual gate. Only then consider merge/unload into the Gamma 30B-A6B candidate.
- Downloads last month
- 14
Model tree for GammaAGI/Gamma-30B-A6B-G1-311-HardNegativeScriptMeta-LoRA
Base model
nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16