FinJev 4B

FinJev-v0 is a financial multiple-choice decision adapter for Qwen3-4B-Instruct-2507. This repository contains the validation-selected epoch-3 LoRA adapter and tokenizer, not merged 4B backbone weights. The pinned base revision is cdbee75f17c01a7cc42f958dc650907174af0554.

The adapter is trained by cross entropy restricted to the supplied option-letter tokens, rather than free-form answer generation. Use the included predict.py for the evaluated scoring method. There is no new JEPA architecture or trained long-reasoning output format in this prototype.

Usage

Install requirements and download this repository, then run:

pip install -r requirements.txt
python predict.py \
  --context 'The company reported higher profit and raised its full-year outlook.' \
  --question 'What is the financial sentiment?' \
  --choices '["negative", "neutral", "positive"]'

The default adapter path is the script's directory. The base model is downloaded separately at its pinned revision. --base-model can point to matching local weights. Input must contain 2–29 choices and a complete prompt of at most 2,048 tokens. The script refuses overflow instead of truncating evidence. Probabilities are normalized over the supplied candidates and are not calibrated confidence estimates. The caller's choice order is retained; benchmark evaluation used a fixed, reproducible permutation per example.

Training

Training data: FPB, FiQA-SA, English six-topic MultiFin, given-pair single-relation FinRED, derived FinQA multiple-choice and arithmetic/count TAT-QA multiple-choice. There are 21,639 eligible training, 3,652 validation and 4,042 test rows. FOMC's 496 test cases were excluded from this checkpoint's training and selection.

LoRA rank 16, alpha 32, dropout 0.05, q/k/v/o projections. Frozen FP16 backbone, FP32 adapter parameters; effective batch 32 on four V100 16GB GPUs. The initial epoch used LR 1e-4; two continuation epochs restored Adam state and restarted LR at 2e-5 with linear decay. Original RNG state was not saved. This is not the same schedule as an uninterrupted three-epoch run. Epoch 3 was selected by the six-task validation mean, before its additional test/FOMC evaluation. Continuation was exploratory after inspection of first-epoch test results.

Results

Accuracy on the same frozen eligible test cases, in percent:

Task	N	Base Qwen 4B	Fin-R1 7B	FinJev 4B
FPB	726	61.02	73.00	88.98
FiQA	234	87.61	84.62	85.90
MultiFin	508	52.95	49.41	84.45
FinRED	1,075	28.09	34.79	83.63
FinQA MC	763	41.42	87.16	88.34
TAT-QA MC	736	47.69	94.02	94.57
Six-task unweighted mean	4,042	53.13	70.50	87.64
FOMC transfer	496	54.23	58.27	59.48