yusifnuri/phi-4-mini-instruct_financial_sentiment

phi-4-mini-instruct — Financial sentiment adapter

A LoRA adapter that specialises microsoft/Phi-4-mini-instruct (3.80 B parameters) for a single enterprise task: it classifies a financial sentence as negative, neutral or positive.

It was produced for the MSc thesis Fine-Tune or Pay Per Token? An Enterprise Benchmark of Small Language Models (SRH University Hamburg), which measures fine-tuned small models against frontier provider APIs on accuracy, latency, cost, privacy exposure and return-on-investment breakeven volume. The adapter is released so that the benchmark can be independently verified.

Read this before using the adapter

• The training corpus (Financial PhraseBank) is licensed CC BY-NC-SA 3.0. This adapter is a research artefact and is not commercially deployable; commercial use would require a licensed corpus or your own annotations.

• This adapter scores below its own split's neutral-majority share (~0.61), which indicates a failure to produce the target output format rather than partial competence. Do not use it. The cause is under investigation; see §4.2.4 of the thesis.

Measured performance

Metric	Value
Accuracy	0.545
Mean latency, batch 1	522 ms
Cost per 1M generated tokens	USD 18.09

Measured on a single NVIDIA H200 (141 GB) at batch size one and full utilisation, priced at an imputed USD 3.99 per GPU-hour. Latency excludes network transit. Scores are not comparable across tasks — each task carries its own metric. Evaluation ran on 5 July 2026; the complete matrix is at results/benchmark_matrix.csv.

Training

Method	LoRA
Dataset	Financial PhraseBank (AllAgree) ((not redistributable))
Dataset licence	CC BY-NC-SA 3.0 — non-commercial
Training examples	5,000 (500 held out for checkpoint selection)
Rank / alpha / dropout	16 / 32 / 0.05
Target modules	q_proj, k_proj, v_proj, o_proj
Learning rate	2e-4, cosine schedule, 3% warmup
Epochs	3
Effective batch size	16 (4 x 4 gradient accumulation)
Max sequence length	512 tokens
Optimiser	AdamW
Seed	42

Hyperparameters were held constant across every model and task rather than tuned per cell, so these figures are a conservative lower bound on attainable performance.

Usage

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained("microsoft/Phi-4-mini-instruct", device_map="auto")
model = PeftModel.from_pretrained(base, "<your-hf-username>/phi-4-mini-instruct_financial_sentiment")
tokenizer = AutoTokenizer.from_pretrained("microsoft/Phi-4-mini-instruct")

The adapter was trained on this prompt format and expects it at inference:

Classify the sentiment of this financial sentence (negative / neutral / positive):
{text}
Sentiment:

Limitations

• Trained once, with a single seed. Reported differences confound model quality with initialisation variance.
• Specialised to one task on one public corpus. It is not a general-purpose assistant and should not be treated as one.
• The evaluation corpora are long-standing public benchmarks and are plausibly present in the base model's pretraining data, which inflates absolute scores.
• Evaluation used 200 held-out instances (all 164 problems for code generation), so detectable effect sizes are bounded at roughly ten percentage points.

Links

• Code, configurations and evaluation harness: https://github.com/Yusifnuri/slm-benchmark
• Full benchmark matrix: results/benchmark_matrix.csv
• Per-request cost analysis: results/cost_per_request.csv

Citation

@mastersthesis{nuri2026finetune,
  title  = {Fine-Tune or Pay Per Token? An Enterprise Benchmark of Small Language Models},
  author = {Nuri, Yusif},
  school = {SRH University Hamburg},
  year   = {2026}
}