Etherlabs/finance-ops-triage-v0.1

Finance Ops Triage v0.1

An educational Unsloth MLX fine-tuning and local deployment experiment by Ugo Chukwu. Deployment succeeded; model quality remains imperfect.

Code and complete experiment report · Dataset

Files and compatibility

• Qwen2.5-0.5B-Instruct.Q4_K_M.gguf: exported local inference model, 397,807,936 bytes (397.81 decimal MB). Its GGUF metadata identifies Qwen2 architecture and file type 15 (Q4_K_M). The earlier UI estimate of roughly 291 MB was not the measured file size.
• adapter/: original Unsloth MLX adapter, tokenizer, chat template, configuration and row-bound seed metadata.
• adapter/checkpoint-21/: final checkpoint, trainer history and optimizer state, retained as experiment evidence. Resume has not been tested.
• model_artifacts.json: SHA-256 hashes and sizes for the original model files.

The saved adapter requires Unsloth MLX runtime quantization. Do not assume standard Transformers/PEFT loading works. The verified deployment path was loading the GGUF in Unsloth and calling its authenticated OpenAI-compatible localhost API. No hosted inference deployment or automatic Hub inference compatibility is claimed.

Training

MacBook Pro M5, 24 GB; Qwen2.5-0.5B-Instruct-bnb-4bit; 50 training examples, 15 validation examples; 3 epochs; 512-token context; learning rate 2e-4; batch 2 × accumulation 4 = 8; 21 optimizer steps. Adapter rank 16, scale 1.0, dropout 0.0; 24 layers; q/k/v/o and gate/up/down projections. Runtime quantization was 4-bit affine, group size 64.

Final saved training loss: 0.1580784768. Validation loss declined from approximately 1.9384 to 0.3987. Loss is not task accuracy.

Evaluation and limitations

Historical manual scoring on five validation cases, one point each for category, severity, single JSON, and sensible action:

Criterion	Base	Fine-tuned
Category	3/5	3/5
Severity	2/5	1/5
Single JSON	2/5	4/5
Next action	1/5	4/5
Composite	8/20 (40%)	12/20 (60%)

These are validation cases, not an untouched test set. The scores were collected before GGUF export; no matched quantization benchmark was run. The fine-tune improved formatting and operational wording while category boundaries remained weak, severity tended toward medium, and one response continued with extra JSON objects. Later GGUF API calls worked but still misclassified missing and duplicate transaction cases.

This is not a production-qualified model, financial policy, or system that can safely change a ledger. Generated actions require review. No robust schema enforcement is built into the original client. No controlled latency benchmark or general reliability claim is made.

Local usage

Download the GGUF, add its folder to Unsloth's on-device Model Hub, and load the model. Keep the local API running. The archived client uses http://127.0.0.1:8888/v1, model ID Qwen2.5-0.5B-Instruct.Q4_K_M, and a server key supplied through UNSLOTH_API_KEY. Follow the GitHub README for setup. No credentials are included.

Provenance and license

Derived from the Qwen2.5-0.5B instruction model, fine-tuned with the associated illustrative Finance Ops Triage dataset and exported through Unsloth. The upstream Apache-2.0 license is included. This repository contains a modified model and records its fine-tuning and quantization here. The original source checkpoints and data remain preserved locally.