mapfinben-qwen3-merged-unified

mapfinben-qwen3-merged-unified

Unified LoRA fine-tuned Qwen3-8B-Instruct for CCL26-Eval-MapFinBen (five financial NLP tasks in one model).

Model Details

• Base model: Qwen3-8B-Instruct
• Method: LoRA SFT (rank=16, alpha=32, 1 epoch)
• Training data: MapFinBen train split, 51,064 samples (AS+QA+SA+TC+TS merged)
• Framework: LLaMA-Factory
• Parameters: ~8B (merged full weights)

Tasks

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model_path = "Ljy2004/mapfinben-qwen3-merged-unified"
tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_path, trust_remote_code=True, torch_dtype="auto", device_map="auto"
)

messages = [{"role": "user", "content": "YOUR_PROMPT"}]
text = tokenizer.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True, enable_thinking=False
)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512)

Important: Use Qwen3 chat template with enable_thinking=False to match training.

Training Hyperparameters

• LoRA rank: 16, alpha: 32, dropout: 0.05, target: all
• Learning rate: 1e-4, cosine scheduler, warmup 3%
• Batch size: 1 × grad_accum 8, cutoff_len: 2048, bf16

Citation

MapFinBen benchmark: https://github.com/MapFinBen/MapFinBen