Amsi-fin-o1.5

This release replaces the previous weak AITRADER finance model line with a full-parameter Qwen 3.5 domain update focused on trading strategy reasoning, market structure interpretation, risk framing, and finance question answering.

What Changed

• Base model: Qwen/Qwen3.5-9B-Base
• Fine-tuning mode: full_finetuning=true
• Optimizer: adafactor
• Learning rate: 1e-05
• Sequence length: 1024
• Gradient accumulation: 4
• Source artifact used for packaging: full
• Packaging date: 2026-03-15

Intended Behavior

• Stronger text-only finance and trading strategy reasoning than the prior AITRADER/Amsi-fin-o1 line.
• Better regime, invalidation, and risk explanation behavior.
• Cleaner supervised tuning footprint for OpenAI-compatible serving and downstream benchmarking.

Local Benchmark Results

Pending. Run bash scripts/run_release_benchmarks.sh after training completes.

Public Comparison References

Model	Benchmark	Public score	Note	Source
AITRADER/Amsi-fin-o1	Hugging Face model card	-	Vision-language 4B predecessor; no public FinEval or BizFinBench score published in the model card.	https://huggingface.co/AITRADER/Amsi-fin-o1
Qwen2.5-7B-Instruct	FinEval text weighted average	69.7	Public FinEval text leaderboard score.	https://github.com/SUFE-AIFLM-Lab/FinEval
GLM-4-9B-Chat	FinEval Chinese financial weighted average	58.4	Public FinEval Chinese financial domain leaderboard score.	https://github.com/SUFE-AIFLM-Lab/FinEval
FinGPTv3.1	FinEval Chinese financial weighted average	27.1	Public FinEval Chinese financial domain leaderboard score.	https://github.com/SUFE-AIFLM-Lab/FinEval
GPT-4o	BizFinBench average	71.8	Public BizFinBench v1 average score.	https://github.com/HiThink-Research/BizFinBench
ChatGPT-o3	BizFinBench average	73.86	Public BizFinBench v1 average score.	https://github.com/HiThink-Research/BizFinBench

These public figures come from external leaderboards and model cards. They are not apples-to-apples with this run unless the same benchmark protocol and prompt format are used. They are included as reference baselines for the later release note.

Training Configuration Snapshot

base_model: Qwen/Qwen3.5-9B-Base
max_seq_length: 1024
full_finetuning: True
gradient_checkpointing: True
learning_rate: 1e-05
weight_decay: 0.01
per_device_train_batch_size: 1
gradient_accumulation_steps: 4
warmup_steps: 100
max_steps: 12000
optim: adafactor
seed: 3407

Release Workflow

1. Finish training and confirm the final checkpoint or full/ export.
2. Run bash scripts/run_release_benchmarks.sh.
3. Run bash scripts/prepare_hf_update.sh.
4. Review the generated bundle under artifacts/releases/.
5. Push the bundle contents to AITRADER/Amsi-fin-o1.5 when satisfied.