darkmem-classifier-v1

Seven-class memory-type classifier for darkmem. Labels: fact, decision, preference, problem, reference, architecture, milestone.

Metrics

accuracy 0.975 / macro F1 0.975 on 1,000-row gold (gold_v3.jsonl).

Base model

answerdotai/ModernBERT-base

Usage

from transformers import AutoModelForSequenceClassification, AutoTokenizer

tok = AutoTokenizer.from_pretrained("darkraise/darkmem-classifier-v1", trust_remote_code=True)
model = AutoModelForSequenceClassification.from_pretrained("darkraise/darkmem-classifier-v1", trust_remote_code=True)

License

Inherits the license of the base model. Fine-tuned weights published under the same terms unless noted otherwise in the repo.

Provenance

Fine-tuned as part of darkmem — a centralized memory system for AI agents. Training recipe and evaluation scripts are in the fine-tuning/ subtree of the darkmem repository.

Downloads last month
37
Safetensors
Model size
0.1B params
Tensor type
F32
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for darkraise/darkmem-classifier-v1

Finetuned
(1249)
this model