license, base_model, datasets, language, pipeline_tag, tags
license base_model datasets language pipeline_tag tags
apache-2.0 mistralai/Mistral-7B-Instruct-v0.3
PolyAI/banking77
en
text-generation
banking77
intent-classification
qlora
bf16

banking77-mistral-7b-v0.3

mistralai/Mistral-7B-Instruct-v0.3 fine-tuned with QLoRA (Unsloth, r = 16, 2 epochs) on the Banking77 train split (10 003 queries) to classify bank customer queries into 77 intents. Part of a master's thesis on local LLM inference for customer support (National Aviation University, Kyiv, 2026).

Format: Merged BF16 weights (LoRA adapter merged into the original 16-bit base model). Serve with vLLM / transformers.

Prompt (single user turn, the model's own chat template); the answer is the intent name:

Classify the bank customer query into one intent. Answer with the intent name only.

Query: <customer query>

Banking77 test accuracy (3 080 queries, NF4 base + adapter, before merging/quantization): 0.9477. Accuracy of this exact format is measured separately.

License: Apache 2.0, as the base model. Banking77 — CC BY 4.0 (Casanueva et al., 2020).

S
Description
Model synced from source: Jorbaty/banking77-mistral-7b-v0.3
Readme
656 KiB
0 Stars 1 Watchers 0 Forks
Languages
Jinja 100%