augmem/ESS-AIST-81M-preview-GGUF

ESS-AIST-81M Preview GGUF

This repository contains GGUF quantizations of augmem/ESS-AIST-81M-preview.

Base model:

• augmem/ESS-AIST-81M-preview

Quantizations:

• ESS-AIST-81M_q8_0.gguf
• ESS-AIST-81M_q5_1.gguf

These files were produced from the current v9 Cortext trial checkpoint:

• source checkpoint: ess_aist_full_v9_subjectfix_l4k/best_model.pt
• exported checkpoint epoch: 3

Exact Release Metrics

The quantized files correspond to the same release checkpoint and eval bundle as the base repo.

All numbers below are from the exact published checkpoint state exported from ess_aist_full_v9_subjectfix_l4k/best_model.pt at checkpoint epoch 3.

Evaluation scope note:

• SALT is train-adjacent for this ESS line because SALT-derived rows were included in training.
• speech holdout is also train-adjacent because explicit speech/audio-text supervision was added back into the corpus.
• Treat these numbers as release regression gates for the source checkpoint, not as contamination-free external benchmark claims.
• A later full external sweep (MTEB / MIEB / MAEB) is still pending.

Speech holdout:

• A->T_r1 = 0.3276
• T->A_r1 = 0.3202

SALT:

• I->T_r1 = 0.3179
• T->I_r1 = 0.3425
• A->T_r1 = 0.1226
• T->A_r1 = 0.1272

Held-out ESS:

• subject_key same/different AUC: 0.9881
• subject_key same-topic-different-subject rejection AUC: 0.9881
• event_key same/different AUC: 0.8193
• subject_key same-subject-different-event rejection AUC: 0.7381

Files

File	Purpose
ESS-AIST-81M_q8_0.gguf	Higher-accuracy GGUF
ESS-AIST-81M_q5_1.gguf	Smaller GGUF
manifest.json	Release manifest
parameter_breakdown.json	Exact parameter accounting
retrieval_512_gt1030.json	Exact 512d retrieval eval for the source checkpoint
subject_eval.json	Exact held-out subject eval for the source checkpoint
event_eval.json	Exact held-out event eval for the source checkpoint
prefix_eval.json	Prefix-level AUC summary

Notes

• This is a preview quantization repo for internal evaluation.
• The source checkpoint is still a bridge artifact under active architecture work.
• This v9 preview materially improves held-out subject/entity separation, but retrieval is weaker than the earlier v7 preview.
• The attached SALT and speech holdout numbers are inherited from the source checkpoint's train-adjacent eval bundle.