XO GPT: Conversation Summarization Model

Back to XO GPT Model Specifications

The XO GPT Conversation Summarization model generates concise, context-aware summaries of agent-customer interactions. It uses abstractive summarization, context analysis, and sentiment detection to transform lengthy dialogues into actionable insights.

---

Challenges with Commercial Models

Challenge	Impact
Latency	High processing times affect user experience in real-time or high-volume scenarios.
Cost	Per-request pricing scales poorly for large deployments.
Data Governance	Sending conversations to external models raises privacy and security concerns.
Lack of Customization	General-purpose models are not tuned for specific industries or use cases.
Limited Control	Minimal ability to correct or refine model behavior for incorrect outputs.
Compliance Constraints	Some industries have regulatory requirements that commercial LLM providers don’t fully support.

---

Key Assumptions

• Designed for text-based conversations only.
• Assumes structured conversational data with clear speaker delineation.

---

Benefits

Consistent and Accurate

Delivers precise, contextually relevant summaries for conversation transcripts. See Model Benchmarks for latency and accuracy metrics.

Cost-Effective

For Enterprise Tier customers, XO GPT eliminates commercial model usage costs. Example comparison (250 input tokens/conversation, 1,000 daily summaries, 120 tokens/summary):

Model	Input $/MTok	Output $/MTok	Input $/Year	Output $/Year	Total $/Year
GPT-4 Turbo	$30	$60	$2,738	$2,628	$5,366
GPT-4	$10	$30	$913	$1,314	$2,227
GPT-4o Mini	$0.15	$0.60	$13.69	$26.28	$39.97

Enhanced Security

No client or user data is used for model retraining.

Guardrails: Content moderation, behavioral guidelines, response oversight, input validation, and usage controls.

AI Safety: Ethical guidelines, bias monitoring, transparency, and continuous improvement.

Performance, features, and language support may vary by implementation. Test thoroughly in your environment before production use.

---

Use Cases

Domain	Use Case
Healthcare	Summarize patient inquiries about symptoms, medications, and follow-up instructions.
Banking	Summarize conversations about account issues, transaction disputes, or loan applications.
E-commerce	Summarize inquiries about product availability, order status, returns, and refunds.
Insurance	Summarize policyholder interactions about claims, policy updates, and coverage questions.
IT Support	Summarize troubleshooting steps, error reports, and resolutions for technical issues.
Telecommunications	Summarize complaints and service requests about network issues, billing errors, and plan changes.
Travel & Hospitality	Summarize queries about booking modifications, cancellations, and special requests.
Retail	Summarize interactions about store policies, promotions, and product exchanges.
Education	Summarize inquiries about course enrollments, schedules, and academic records.
Utilities	Summarize communications about service outages, bill inquiries, and usage reports.

---

Sample Output

Conversation:

App: Hello! How can I help you today?
Customer: I need to check the status of my order.
App: Sure! Please provide your order reference number.
Customer: It's 12345-67890.
[Identity verified; order confirmed shipping within 48 hours]
Customer: Yes, I want to speak with an agent.
Agent: Hi, this is John from XYZ Support. How can I assist you today?
Customer: I wanted to confirm the shipping address on my order.
Agent: The address on file is 123 Elm Street, Springfield, IL.
Customer: That's correct. Thanks!

Generated Summary:

The customer contacted support to check the status of their order. The AI Agent verified the customer’s identity and informed them their order would ship within 48 hours. The customer then requested to speak with an agent to verify their shipping address. The agent confirmed the address on file was correct. The conversation ended with the customer satisfied.

---

Model Building Process

See Model Building Process.

---

Model Benchmarks

Version	Accuracy	TPS	Latency (s)	Benchmark	Test Data
v2.0	100%	71	2.00	Summary v2	Results v2
v1.0	98%	40	3.04	Summary v1	Results v1

---

Version 2.0

Model Choice

Base model: Mistral 7B Instruct v0.2

Base Model	Developer	Language	Release Date	Status	Knowledge Cutoff
Mistral 7B Instruct v0.2	Mistral AI	Multi-lingual	September 2024	Static	September 2024

Fine-Tuning Parameters

Parameter	Description	Value
Fine-Tuning Type	Method used	peft-qlora
Quantization	Bits for loading parameters	4-bit
Rank	Number of trainable parameters	32
LoRA Dropout	Prevents co-adaptation in neural network	0.05
LoRA Alpha	Scaling factor	—
Learning Rate	Rate toward loss minimum	2e-4 (0.0002)
Batch Size	Examples per training step	2
Epochs	Passes over training data	3
Max Sequence Length	Maximum input length	32768
Optimizer	Optimization algorithm	paged_adamw_8bit
Task Type	LoRA task type	CAUSAL_LM
Targeted Modules	Layers where LoRA is applied	["up_proj", "o_proj", "down_proj", "gate_proj", "q_proj", "k_proj", "v_proj"]

General Parameters

Infrastructure: 2× A10 GPUs. Requires an Agent AI License.

Parameter	Description	Value
Learning Rate	Rate toward loss minimum	2e-4 (0.0002)
Batch Size	Examples per training step	2
Epochs	Passes over training data	3
Max Sequence Length	Maximum input length	32768
Optimizer	Optimization algorithm	paged_adamw_8bit

AWQ Model Quantization

Parameter	Description	Value
Zero Point	Include zero-point for better weight representation	True
Quantization Group Size	Weight group size	128
Weight Precision	Bits for weight representation	4
Quantization Version	AWQ version for GEMM operations	”GEMM”
Computation Data Type	Data type for inference	torch.float16
Model Loading	Reduced CPU memory usage	{"low_cpu_mem_usage": True}
Tokenizer Loading	Remote code compatibility	trust_remote_code=True

Benchmarks Summary v2

Comparison models: LLama-8B, GPT-4, Claude 3 Sonnet.

XO GPT achieved an outstanding overall score, positioning it alongside Llama and ahead of Sonnet and GPT-4. It delivers strong results across accuracy, fluency, and robustness in English, French, German, Japanese, Turkish, and Spanish.

See Test Data and Results v2 for full details.

---

Version 1.0

Model Choice

Base model: Mistral 7B Instruct v0.2

Base Model	Developer	Language	Release Date	Status	Knowledge Cutoff
Mistral 7B Instruct v0.2	Mistral AI	Multi-lingual	March 2024	Static	September 2024

Fine-Tuning Parameters

Parameter	Description	Value
Fine-Tuning Type	Method used	peft-qlora
Quantization	Bits for loading parameters	4-bit
Rank	Number of trainable parameters	32
LoRA Dropout	Prevents co-adaptation in neural network	0.05
Learning Rate	Rate toward loss minimum	1e-3 (0.001)
Batch Size	Examples per training step	2
Epochs	Passes over training data	4
Max Sequence Length	Maximum input length	—
Optimizer	Optimization algorithm	paged_adamw_8bit
Task Type	LoRA task type	CAUSAL_LM

General Parameters

Infrastructure: 2× A10 GPUs. Requires an Agent AI License.

Parameter	Description	Value
Learning Rate	Rate toward loss minimum	1e-3 (0.001)
Batch Size	Examples per training step	2
Epochs	Passes over training data	4
Max Sequence Length	Maximum input length	32768
Optimizer	Optimization algorithm	paged_adamw_8bit

Benchmarks Summary v1

Comparison models: Llama 3 8B (Ctranslate), Sonnet 3.5, GPT-4o.

XO GPT demonstrates strong performance in English, French, German, and Spanish, with notable results in bias detection, sentiment analysis, and negation detection.

See Test Data and Results v1 for full details.