Training lineage
The upstream model was first fine-tuned in run
73db357b-2052-45f2-ae46-55db66c4bebf (fp32_mixed_task). Its merged
checkpoint was then used as the base for run
097f068b-a3d0-4419-bd18-05044b1d72af, named
full_1000_from_73db_2026-10-06. The final adapters were merged into that
intermediate checkpoint. The base_model metadata names the upstream Hub
model; the intermediate checkpoint is a local training artifact.
- Full merged weights in
model.safetensors.
- The exported model configuration declares
float32 precision.
- Matching tokenizer, generation configuration and chat template are included.
- This folder contains the full-precision checkpoint preceding the Astra
quantization, pruning and LiteRT-LM graph changes.
Scope and evaluation
The model was fine-tuned for the project's game-specific tasks. Use the
matching task prompts, output schema and chat template. No new inference or
evaluation was run while preparing this model card. Evaluation results for
quantized or pruned Astra exports should be attributed to those exports.
License
The upstream Gemma 4 model is distributed under Apache License 2.0.
Preserve the applicable upstream license and attribution files when sharing
this checkpoint.