yusifnuri/Mistral-7B-v0.3_classification

Mistral-7B-v0.3 — Topic classification adapter

A QLoRA (4-bit NF4, double quantisation) adapter that specialises mistralai/Mistral-7B-v0.3 (7.25 B parameters) for a single enterprise task: it assigns a news item to one of four topics (World, Sports, Business, Sci/Tech).

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

• Adapted with QLoRA from the base release, not an instruction-tuned one. Any deficit is jointly attributable to the model and to 4-bit adaptation; the two cannot be separated within this design.

Measured performance

Metric	Value
Accuracy	0.69
Mean latency, batch 1	556 ms
Cost per 1M generated tokens	USD 19.25

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	QLoRA (4-bit NF4, double quantisation)
Dataset	AG News (fancyzhx/ag_news)
Dataset licence	Custom, research use
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 (2 x 8 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("mistralai/Mistral-7B-v0.3", device_map="auto")
model = PeftModel.from_pretrained(base, "<your-hf-username>/Mistral-7B-v0.3_classification")
tokenizer = AutoTokenizer.from_pretrained("mistralai/Mistral-7B-v0.3")

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

Classify the following news text into exactly one category (World / Sports / Business / Technology):
{text}
Category:

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}
}