sll0107/Qwen3-8B-finance

Qwen3-8B-finance

Qwen3-8B fine-tuned for multilingual financial NLP (sentiment analysis, topic classification, question answering, summarization, and multiple-choice).

Inference notes

• —Decoding: greedy (temperature=0, top_p=1).
• —Prompt format: the model is trained to receive the prompt verbatim — no system prompt and no chat wrapping. The embedded chat_template (in tokenizer_config.json) is intentionally pass-through for this reason.

Running with ollama (HF → GGUF → ollama)

Convert to GGUF (f16), then create the ollama model with a pass-through template so the prompt is fed as-is (this avoids a default chat template reshaping the input):

# Modelfile
FROM ./model-f16.gguf
PARAMETER temperature 0
PARAMETER top_p 1
PARAMETER seed 42
PARAMETER stop "<|im_end|>"
PARAMETER stop "<|endoftext|>"
TEMPLATE """{{ .Prompt }}"""

The same snippet is provided in template.tmpl.