andreeaduti's picture
Add fine-tuned decoder (mistral_7b)
febf3ce verified
|
Raw
History Blame Contribute Delete
789 Bytes
metadata
license: apache-2.0
base_model: mistralai/Mistral-7B-Instruct-v0.3
tags:
  - cyber-threat-intelligence
  - information-extraction

decoder (mistral_7b)

Fine-tuned mistralai/Mistral-7B-Instruct-v0.3 for structured vulnerability-record extraction from cybersecurity news articles, from the study (encoder token-classification vs. decoder structured-generation). This is the best-performing decoder in that comparison.

Test-set performance (entity/field-level)

parse rate micro F1 macro F1
0.9423 0.5401 0.5852

See the project README for the task definition, schema, evaluation protocol, and the full per-model / per-field breakdown.

Companion dataset: upb-nlp/NtVR.