SetFit with lightonai/modernbert-embed-large

This is a SetFit model that can be used for Text Classification. This SetFit model uses lightonai/modernbert-embed-large as the Sentence Transformer embedding model. A LogisticRegression instance is used for classification.

The model has been trained using an efficient few-shot learning technique that involves:

Fine-tuning a Sentence Transformer with contrastive learning.
Training a classification head with features from the fine-tuned Sentence Transformer.

Model Details

Model Description

Model Type: SetFit
Sentence Transformer body: lightonai/modernbert-embed-large
Classification head: a LogisticRegression instance
Maximum Sequence Length: 8192 tokens
Number of Classes: 6 classes

Model Sources

Repository: SetFit on GitHub
Paper: Efficient Few-Shot Learning Without Prompts
Blogpost: SetFit: Efficient Few-Shot Learning Without Prompts

Model Labels

Label	Examples
1	'what can your do with your mcp tool db2 schema' 'is this normal?\nGC(47) Pause Young 234M->89M 12.5ms' 'what units is file_size in? what format is mimetype in'
0	'use mongo and postgres mcps to find test data' 'browser screenshot the error page' 'use the linear mcp to find my high priority issues'
2	'java\npublic void process() {\n if (user != null) {\n // stuff\n }\n}\n\nclean this up' "Duplicate class: 'ChunkCacheMetricsTest'" "python\ndef process(data):\n result = []\n for item in data:\n if item['active']:\n result.append(item['value'])\n return result\n\nuse list comprehension"
3	'review security scan then patch critical issues' 'analyze failing tests, group them, fix each group' '\n47 tests failing\n\ncategorize and fix each category'
5	'model architecture not supported, plan alternative approach' 'architecture diagram for wml' 'before we move on let me ask - can we develop a set of documents that break the implementation down into individual tasks - the first one being a simple core set of function that could deliver a a single starting point. and then add layers of function in subsequent tasks to methodically build out th'
4	'yes the linear mcp is finally connected' 'Usually we use a custom linter config here, not the standard one' 'ok it worked after i restarted the postgres mcp server'

Uses

Direct Use for Inference

First install the SetFit library:

pip install setfit

Then you can load this model and run inference.

from setfit import SetFitModel

# Download from the 🤗 Hub
model = SetFitModel.from_pretrained("tmp/best_model")
# Run inference
preds = model("lets fix the issues")

Training Details

Training Set Metrics

Training set	Min	Median	Max
Word count	1	12.1876	125

Label	Training Sample Count
0	43
1	80
2	92
3	56
4	64
5	86

Training Hyperparameters

batch_size: (8, 8)
num_epochs: (2, 2)
max_steps: -1
sampling_strategy: undersampling
num_iterations: 10
body_learning_rate: (1e-05, 1e-05)
head_learning_rate: 0.0011800069021089953
loss: CosineSimilarityLoss
distance_metric: cosine_distance
margin: 0.25
end_to_end: False
use_amp: False
warmup_proportion: 0.006100049987809886
l2_weight: 0.01
seed: 42
eval_max_steps: -1
load_best_model_at_end: True

Training Results

Epoch	Step	Training Loss	Validation Loss
0.0009	1	0.2413	-
0.0199	21	-	0.2284
0.0399	42	-	0.2249
0.0475	50	0.2391	-
0.0598	63	-	0.2136
0.0798	84	-	0.1990
0.0950	100	0.2046	-
0.0997	105	-	0.1808
0.1197	126	-	0.1670
0.1396	147	-	0.1388
0.1425	150	0.1388	-
0.1595	168	-	0.1421
0.1795	189	-	0.1296
0.1899	200	0.0858	-
0.1994	210	-	0.1442
0.2194	231	-	0.1470
0.2374	250	0.0469	-
0.2393	252	-	0.1401
0.2593	273	-	0.1462
0.2792	294	-	0.1243
0.2849	300	0.0272	-
0.2991	315	-	0.1160
0.3191	336	-	0.1211
0.3324	350	0.0126	-
0.3390	357	-	0.1028
0.3590	378	-	0.1212
0.3789	399	-	0.1173
0.3799	400	0.0075	-
0.3989	420	-	0.1180
0.4188	441	-	0.1269
0.4274	450	0.0071	-
0.4387	462	-	0.1306

Framework Versions

Python: 3.12.12
SetFit: 1.1.3
Sentence Transformers: 3.4.1
Transformers: 4.51.3
PyTorch: 2.11.0+cu130
Datasets: 4.8.4
Tokenizers: 0.21.4

Citation

BibTeX

@article{https://doi.org/10.48550/arxiv.2209.11055,
    doi = {10.48550/ARXIV.2209.11055},
    url = {https://arxiv.org/abs/2209.11055},
    author = {Tunstall, Lewis and Reimers, Nils and Jo, Unso Eun Seo and Bates, Luke and Korat, Daniel and Wasserblat, Moshe and Pereg, Oren},
    keywords = {Computation and Language (cs.CL), FOS: Computer and information sciences, FOS: Computer and information sciences},
    title = {Efficient Few-Shot Learning Without Prompts},
    publisher = {arXiv},
    year = {2022},
    copyright = {Creative Commons Attribution 4.0 International}
}