CAI-20B-v2-GGUF

Marketing-specialized 20B parameter LLM fine-tuned on proprietary Meta ads, scaling, and creative optimization frameworks.

Model Details

Attribute	Value
Base Model	GPT-OSS-20B
Architecture	Mixture-of-Experts (32 experts, 4 active)
Context Length	131,072 tokens
Training	LoRA fine-tuned on 588 curated marketing Q&A pairs
License	MIT

Available Quantizations

All quantizations use imatrix (importance matrix) computed from domain-specific marketing calibration data for optimal quality preservation.

Quantization	Size	Description	Use Case
Q8_0	21 GB	8-bit quantization	Maximum quality, requires 24GB+ VRAM
Q5_K_M	16 GB	5-bit mixed precision	Best balance of quality/size
Q4_K_M	15 GB	4-bit mixed precision	Good quality, moderate VRAM
Q4_K_S	14 GB	4-bit small	Minimum VRAM, still usable

imatrix Calibration

The importance matrix was computed using marketing-specific calibration data:

• Meta ads terminology and frameworks
• Scaling methodologies (Chad Scaling, Blitz Scaling)
• Creative optimization strategies
• Performance metrics (ROAS, MER, CPA)

Perplexity during imatrix computation: 2.0859

Specializations

• Meta Ads: Campaign structure, ABO vs CBO, audience targeting, budget optimization
• Scaling Frameworks: Chad Scaling, Blitz Scaling, systematic growth methodologies
• Creative Strategy: Hook frameworks, UGC structure, creative fatigue solutions
• Performance Metrics: ROAS optimization, MER targeting, attribution modeling

---

Evaluation Results

LLM-as-Judge Benchmark (Grok 4.1 Thinking)

Rigorous pairwise evaluation comparing CAI-20B-v2 (fine-tuned) vs GPT-OSS-20B (base model) using position debiasing.

Test Configuration:

• Judge Model: grok-4-1-fast-reasoning (xAI)
• Quantization Tested: Q4_K_M via Modal serverless inference
• Position Debiasing: Both orderings tested (A,B) and (B,A)
• Prompts: 10 domain-specific marketing questions

Summary Results

Metric	Value
CAI-20B-v2 Win Rate	70%
GPT-OSS-20B Win Rate	30%
Ties	0%
Confident Judgments	100%

Per-Question Breakdown

Question	Winner	Key Finding
Scale FB ads $500→$5000/day	Baseline	Finetuned had repetition issues
Campaign structure for new brand	Finetuned	Specific budget splits ($1.5k/$2k/$1.5k)
Fix creative fatigue	Finetuned	12 specific tactics vs generic guide
Metrics beyond ROAS	Baseline	Finetuned produced video hooks instead
Cold traffic ad copy	Finetuned	Clear Hook/Pain/Solution/Proof/CTA framework
ABO vs CBO decision	Finetuned	Precise decision trees with examples
Finding winning audiences	Finetuned	10-step process vs off-topic response
Systematic creative testing	Finetuned	8-step methodology with A/B testing specifics
Landing page conversion	Finetuned	13 tactics with concrete examples
New product launch playbook	Baseline	Finetuned produced sales copy

Evaluation Criteria (Rubric)

Each response judged on 5 criteria weighted equally:

1. Actionability - Specific, implementable steps
2. Depth - Expert-level understanding
3. Structure - Organization and readability
4. Specificity - Numbers, metrics, examples
5. Practicality - Realistic for real businesses

Key Findings

Strengths of CAI-20B-v2:

• Superior framework-based thinking (Hook/Pain/Solution/Proof/CTA)
• More specific budget allocations and metrics
• Platform-specific tactics (Meta's Dynamic Creative, frequency caps)
• Better decision frameworks for common choices (ABO vs CBO)

Areas for Improvement:

• Occasional repetition loops in some responses
• Some prompts trigger promotional/video-script style outputs
• May benefit from better system prompting for consistency

Evaluation Methodology

Position Debiasing Protocol:
1. Run judgment with (Finetuned=A, Baseline=B)
2. Run judgment with (Baseline=A, Finetuned=B)
3. If both agree → confident result
4. If disagree → mark as TIE (position bias detected)

This eliminates the common LLM-as-judge bias of preferring the first or second response regardless of quality.

---

Usage

LM Studio

1. Download any GGUF file
2. Load in LM Studio
3. Set context length to 4096+ for best results
4. Recommended temperature: 0.7

Ollama

ollama run hf.co/tigres2526/CAI-20B-v2-GGUF:Q4_K_M

llama.cpp

./llama-cli -m CAI-20B-v2-Q4_K_M.gguf -p "How do I scale Facebook ads?" -n 512

Training Data

Fine-tuned on 588 curated Q&A pairs from proprietary marketing frameworks:

• Meta Ads Mastery (~150 examples)
• Blitz Scaling Framework (~140 examples)
• Creative Command Center (~160 examples)
• Iron Media Thesis (~138 examples)

Total training tokens: ~339,276

Citation

@misc{cai20bv2,
  title={CAI-20B-v2: Marketing-Specialized Language Model},
  author={Tigres2526},
  year={2024},
  publisher={HuggingFace},
  url={https://huggingface.co/tigres2526/CAI-20B-v2-GGUF}
}

License

MIT License - Free for commercial and non-commercial use.