Netta1994/setfit_baai_cybereason_gpt-4o_improved-cot-instructions_two_reasoning_only_reasoning_

SetFit with BAAI/bge-base-en-v1.5

This is a SetFit model that can be used for Text Classification. This SetFit model uses BAAI/bge-base-en-v1.5 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:

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

Model Details

Model Description

• Model Type: SetFit
• Sentence Transformer body: BAAI/bge-base-en-v1.5
• Classification head: a LogisticRegression instance
• Maximum Sequence Length: 512 tokens
• Number of Classes: 2 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
0	'Reasoning why the answer may be good:\n- The answer attempts to address the significance of considering all the answers together, which is the core of the question.\n\nReasoning why the answer may be bad:\n- The provided answer does not specifically mention or directly link to the key points raised in the provided document.\n- It lacks specificity and grounding in the text, failing to reference important considerations such as the significance of the machines involved, the severity of the behavior, and the significance of the users involved.\n- It does not directly explain why combining all the answers aids in determining if the behavior is malicious.\n\nFinal Result: Bad' 'The answer provided is:\n\n"The information provided doesn't cover this specific query. Please refer to additional sources or provide more context to obtain a comprehensive answer."\n\n### Evaluation\n\n1. Context Grounding: \n- Negative: The provided document actually contains a detailed procedure for excluding MalOps. The answer incorrectly claims the information does not cover the query.\n\n2. Relevance: \n- Negative: The answer does not address the question of what the process is, instead, it diverts the user to additional sources.\n\n3. Conciseness: \n- Negative: It doesn't provide concise information pertinent to the question asked.\n\n4. Specificity: \n- Negative: The answer is too general and lacks any specifics related to the step-by-step process detailed in the document.\n\n5. Correct Key/Value/Event Name:\n- Negative: No keys or values related to the exclusion process (like “Exclude” or “allowlist”) are mentioned, which the document clearly outlines.\n\n### Reasoning\n\n- Good points: There are actually no good points since the answer is incorrect and not useful.\n- Bad points: The answer dismisses the information when the procedure is clearly mapped out in the document. It fails to provide the necessary details from the document.\n\n### Final Result\nBad' 'Reasoning:\n\nGood Elements:\n1. Conciseness: The provided answer is concise and directly specifies the action needed for a quarantined file.\n2. Relevance: The answer directly addresses the question, mentioning that the file should be un-quarantined.\n\nBad Elements:\n1. Context Grounding: The answer only partially references the document without showing a comprehensive understanding.\n2. Completeness: The answer mentions un-quarantining but does not provide specifics about the subsequent steps needed for submitting a quarantined file as a false positive.\n3. Format and Clarity: It omits the mention of the prerequisite or steps for un-quarantining mentioned in the document.\n\nFinal Result:\nBad \nThe answer is accurate but lacks the necessary depth and context to be fully correct. It should have included the pre-requisite and clarified the steps for submission after un-quarantining.'
1	'Reasoning:\n\nPositive Aspects:\n1. Context Grounding: The answer mentions a dump file and relates it to the contents of the sensor’s RAM at the time of failure, which is consistent with the provided document.\n2. Relevance: The answer is directed at what the computer generates in case of a system failure.\n\nNegative Aspects:\n1. Specificity: While it correctly mentions that the dump file contains the entire contents of the sensor's RAM at the time of the failure, it does not clearly state specifically that it is a "memory dump file," which would be more precise.\n2. Terminology: The document uses the term "memory dump file," which is more accurate and would align better with the terminology used in the question and document.\n3. Conciseness: It could be perceived as slightly verbose given the requirement for direct and precise information.\n\nGiven these points, the answer could be clearer and more specific in the terminology used.\n\nFinal Result: Bad' '**Evaluation Reasoning:**\n\n1. **Context Grounding:**\n - Good Aspect: The answer is derived from the document where it mentions threat detection as a core capability of the platforms to identify cyber security threats.\n - Bad Aspect: The answer is vague and doesn’t leverage the detailed explanation provided in the document, which talks about using advanced engines, AI, ML, and behavioral analysis to identify cyber security threats.\n\n2. **Relevance:**\n - Good Aspect: The answer addresses the specific question about the purpose of the platforms threat detection abilities.\n - Bad Aspect: It fails to incorporate how the document details the processes (use of AI, ML, etc.) to identify threats.\n\n3. Conciseness:\n - Good Aspect: The answer is definitely concise.\n - Bad Aspect: It is overly concise to the point of lacking necessary detail, which makes it too general and uninformative.\n\n4. Specificity:\n - Good Aspect: The basic idea of identifying cyber security threats is correct.\n - Bad Aspect: It omits specific key elements such as AI, machine learning, and behavioral analysis mentioned in the document.\n\n5. Accuracy regarding key/value/event name: \n - Good Aspect: "Identifying cyber security threats" is a correct general purpose.\n - Bad Aspect: The answer does not mention specifically the details related to the ENGINE or specific methods/tools involved, nor does it leverage the details of how the platform operates.\n\nFinal Evaluation Result:\nBad' 'Reasoning:\n\nGood Points:\n- The answer directly addresses the lack of coverage in the provided document regarding the specific query.\n\nBad Points:\n- The answer is incorrect because it states there is no information on the fifth scenario, but the document provides examples, and the evaluator should infer which of the mentioned scenarios could be considered the fifth if it was in sequence.\n- The answer does not leverage any context from the document, failing at providing even an approximate related response.\n- The answer lacks specificity and does not attempt to relate to any scenarios described.\n\nFinal Result: Bad'

Evaluation

Metrics

Label	Accuracy
all	0.4789

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("Netta1994/setfit_baai_cybereason_gpt-4o_improved-cot-instructions_two_reasoning_only_reasoning_")
# Run inference
preds = model("Reasoning why the answer may be good:
- The answer provides a specific URL, which is required by the question.
- It appears to be in the format expected for image URLs as hinted at in the document.

Reasoning why the answer may be bad:
- The provided answer does not match the precise URL given in the document.
- The correct URL for the second query should be `..\/..\/_images\/hunting_http://miller.co`, while the answer contains `hunting_http://www.flores.net/`, which is not mentioned in the document.
- The answer does not reflect careful cross-referencing with the provided document.

Final result: Bad")

Training Details

Training Set Metrics

Training set	Min	Median	Max
Word count	60	128.2029	239

Label	Training Sample Count
0	34
1	35

Training Hyperparameters

• batch_size: (16, 16)
• num_epochs: (5, 5)
• max_steps: -1
• sampling_strategy: oversampling
• num_iterations: 20
• body_learning_rate: (2e-05, 2e-05)
• head_learning_rate: 2e-05
• loss: CosineSimilarityLoss
• distance_metric: cosine_distance
• margin: 0.25
• end_to_end: False
• use_amp: False
• warmup_proportion: 0.1
• l2_weight: 0.01
• seed: 42
• eval_max_steps: -1
• load_best_model_at_end: False

Training Results

Epoch	Step	Training Loss	Validation Loss
0.0058	1	0.2486	-
0.2890	50	0.2626	-
0.5780	100	0.2394	-
0.8671	150	0.1005	-
1.1561	200	0.0028	-
1.4451	250	0.002	-
1.7341	300	0.0018	-
2.0231	350	0.0016	-
2.3121	400	0.0016	-
2.6012	450	0.0014	-
2.8902	500	0.0013	-
3.1792	550	0.0012	-
3.4682	600	0.0012	-
3.7572	650	0.0012	-
4.0462	700	0.0012	-
4.3353	750	0.0012	-
4.6243	800	0.0011	-
4.9133	850	0.0011	-

Framework Versions

• Python: 3.10.14
• SetFit: 1.1.0
• Sentence Transformers: 3.1.0
• Transformers: 4.44.0
• PyTorch: 2.4.1+cu121
• Datasets: 2.19.2
• Tokenizers: 0.19.1

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}
}