{eliza_short_name} ({quant})

{quant_short_name} weights for the {base_eliza_repo_id} full-precision Eliza fine-tune.

What is this

{quant_blurb}

field	value
Quantization scheme	{quant_scheme_name}
Bit-width (weights)	{quant_bits_weights}
Bit-width (KV cache)	{quant_bits_kv}
Reference paper	{quant_paper}
Sibling base repo	{base_eliza_repo_id}
Recommended runtime	{quant_runtime}
Approximate file size	{quant_file_size}
Target hardware	{quant_target_hw}
Expected quality regression vs base	{quant_quality_delta}

Inference

{quant_inference_block}

Provenance

• Source bf16 weights: {base_eliza_repo_id}
• Calibration data (where applicable): held-out validation split of elizaos/eliza-native-v1-sft.
• Training pipeline + quantization scripts: elizaos/eliza-1-training (scripts/quantization/).

Evaluation

Evaluated against the base bf16 sibling. Numbers below are absolute scores on the eliza-1 evaluation suite; subtract from the base card's scores for the regression delta.

{eval_table}

License

Apache-2.0, inherited from the base checkpoint recorded in the sibling model metadata. Same terms as the {base_eliza_repo_id} sibling.