Qwen3.5_4B_ADS

Qwen3.5_4B_ADS is a full fine-tune of unsloth/Qwen3.5-4B for English-to-Korean financial translation. It was trained with the DQS low-QE curriculum using seed 42.

This repository contains both the original Transformers checkpoint and one LM Studio/llama.cpp export:

model.safetensors: original BF16 fine-tuned checkpoint
Qwen3.5_4B_ADS-Q4_K_M.gguf: the only quantized main-model variant
mmproj-Qwen3.5_4B_ADS-BF16.gguf: vision encoder/projector companion

The BF16 mmproj is not an additional LLM quantization variant. It is kept at BF16 for multimodal compatibility and quality.

LM Studio

Use the latest LM Studio runtime and download the Q4_K_M variant:

lms get https://huggingface.co/alwaysgood/Qwen3.5_4B_ADS@Q4_K_M

The matching mmproj file enables image input. For translation, disable thinking and ask for translation-only output, for example:

text

Translate the following English financial text into Korean. Return only the translation.

<source text>

Training and provenance

Tuning: full-parameter supervised fine-tuning
Seed: 42
Selection: low quality-estimation score first (qe_selection_order=low)
Base model thinking during training/evaluation: disabled
Vision layers: not trained; the base model's multimodal components were preserved
Run artifacts: qwen35_4b_it_full_lowqe_seed42
Source revision: fa8166a883d96460cc285b46d66b74a074b4b8d4

The checkpoint does not contain the optional Qwen3.5 MTP auxiliary head even though the inherited config declares one. The GGUF export therefore excludes MTP. Normal autoregressive generation is unaffected; MTP speculative decoding is not provided.

Evaluation

The following scores are from the original BF16 final checkpoint on the 500-row held-out test set. They are not claimed as a separate Q4_K_M evaluation.

Metric	Score
BLEU	27.4731
chrF	47.9918
COMET (`wmt22-comet-da`)	0.8906
COMETKiwi (`wmt22-cometkiwi-da`)	0.8565
XCOMET-XXL	0.8675
MetricX-24 Hybrid XXL (lower is better)	3.8712

Full evaluation records and configuration are available in the linked run.

License and data note

The model weights follow the Apache-2.0 license of the base model. The training corpus aggregates sources with mixed upstream terms; the dataset card is marked license: other. Users are responsible for reviewing the source-specific terms described in alwaysgood/financial-english-source-corpus.