---
license: apache-2.0
language:
- en
base_model:
- mistralai/Mixtral-8x7B-Instruct-v0.1
pipeline_tag: text-generation
library_name: transformers
datasets:
- suehuynh/marketing-instruct-8k
metrics:
- win_rate
tags:
- autoscientist
- marketing
- instruction-tuning
- mixtral
- copywriting
- advertising
- email-marketing
- product-description
- brandvoice
new_version: suehuynh/Marketing-Mixtral-8x7B-v3
---

# Marketing-Mixtral-8x7B

Fine-tuned Mixtral-8x7B-Instruct-v0.1 on a curated marketing instruction 
dataset using AutoScientist for the AutoScientist Challenge 2026 — Marketing Category.

## Model Details

### Model Description

Marketing-Mixtral-8x7B-v2 is a LoRA fine-tuned version of Mixtral-8x7B-Instruct-v0.1,
adapted for marketing copy generation across five task types: ad/social copy,
email marketing, product descriptions, brand voice rewriting, and campaign
data insights.

**Key contribution:** First open marketing LLM optimized for attribute
faithfulness — product description outputs are grounded strictly in provided
specifications, reducing hallucinated features and invented claims.

- **Developed by:** Sue Huynh
- **Model type:** Causal LM, LoRA fine-tuned
- **Language:** English
- **License:** Apache 2.0
- **Finetuned from:** mistralai/Mixtral-8x7B-Instruct-v0.1
- **Training platform:** AutoScientist by Adaption Labs

### Model Sources

- **Dataset:** suehuynh/marketing-instruct-8k
- **Demo:** suehuynh/marketing-mixtral-demo (HuggingFace Spaces)
- **Challenge:** AutoScientist Challenge 2026 — adaptionlabs.ai

## Uses

### Direct Use

Generate marketing copy from structured briefs across five task types:

- **Ad/Social copy:** platform-specific ads with audience and CTA constraints
- **Email marketing:** campaign, lifecycle, and transactional emails with
  subject + preview + body format
- **Product descriptions:** grounded in provided attribute lists only
- **Brand voice rewriting:** neutral copy rewritten to match a specified voice
- **Table-to-insight:** campaign performance data summarized with recommendations

### Downstream Use

Can be fine-tuned further on domain-specific marketing data (e-commerce,
B2B SaaS, local business) for specialized copy generation pipelines.

### Out-of-Scope Use

- Not intended for strategy generation, market research, or competitive analysis
- Not suitable for medical, legal, or financial copywriting without additional fine-tuning
- Product descriptions should always be reviewed for factual accuracy before publication

## Bias, Risks, and Limitations

- Training data is primarily English-language and Western marketing contexts;
  performance may degrade on non-English or culturally specific briefs
- Despite faithfulness training, product descriptions should be human-reviewed
  before publication
- The model may reflect biases present in marketing copy (demographic targeting
  assumptions, persuasion patterns)
- Generated copy should not be published without editorial review

### Recommendations

Always review generated copy before publication. For product descriptions,
verify all claims against the original product specification. Do not use
generated copy as a substitute for human creative judgment.

## How to Get Started with the Model

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "suehuynh/Marketing-Mixtral-8x7B-v2"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.float16,
    device_map="auto",
    load_in_4bit=True,
)

prompt = """Write a product description for an online store listing.
Use ONLY the information provided below — do not invent features,
specifications, or claims.

Product: Wireless Headphones Model-X

Attributes:
- type: over-ear, ANC
- battery: up to 30 hrs (ANC on)
- charging: USB-C, 10 min = 3 hrs
- weight: 250g
- folds_flat: yes, case included

Target length: 100 words. Tone: practical, benefit-led."""

inputs = tokenizer.apply_chat_template(
    [{"role": "user", "content": prompt}],
    tokenize=True,
    add_generation_prompt=True,
    return_tensors="pt"
).to(model.device)

with torch.no_grad():
    output = model.generate(
        inputs,
        max_new_tokens=200,
        do_sample=False,
        pad_token_id=tokenizer.pad_token_id,
    )

new_tokens = output[0][inputs.shape[1]:]
print(tokenizer.decode(new_tokens, skip_special_tokens=True))
```

## Training Details

### Training Data

8,000 examples across 5 marketing task types, assembled from:
- AdaptData-adapted public marketing datasets (ad copy, email, product descriptions)
- High-quality synthetic examples (brand voice, table insights, email)
- Domain augmentation and general-purpose diversity data via AutoScientist

See full data documentation: [suehuynh/marketing-instruct-8k](https://huggingface.co/datasets/suehuynh/marketing-instruct-8k)

### Training Procedure

#### Preprocessing

- Unified instruction schema: `{instruction, input, output}`
- Faithfulness filtering on product descriptions: outputs verified to
  contain only claims grounded in provided attributes
- Deduplication on output hash
- Balanced sampling across task types before final merge

#### Training Hyperparameters

- **Training regime:** bf16 mixed precision
- **Algorithm:** LoRA (Supervised Fine-Tuning)
- **LoRA rank:** 32
- **LoRA alpha:** 64
- **LoRA dropout:** 0
- **Target modules:** all-linear
- **Learning rate:** 1e-4
- **LR scheduler:** cosine with warmup ratio 0.1
- **Epochs:** 1
- **Batch size:** max
- **Gradient clipping:** 1
- **Train on inputs:** false


#### Speeds, Sizes, Times

- **Training platform:** AutoScientist by Adaption Labs
- **Base model size:** 46.7B parameters (Mixtral MoE, ~13B active per token)
- **Training duration:** approximately 1 hour on AutoScientist infrastructure

## Evaluation

### Testing Data, Factors & Metrics

#### Testing Data

Evaluated on AutoScientist's held-out internal marketing test set
(not publicly available) covering diverse marketing copy generation tasks.

#### Metrics

**Win rate:** pairwise LLM-as-judge comparison between finetuned model
output and base Mixtral-8x7B-Instruct output on identical prompts.

### Results

| Metric | Base Mixtral | Finetuned | Delta |
|---|---|---|---|
| Win rate (AS eval) | 17% | 83% | +388% relative |

#### Summary

The finetuned model achieves a 83% win rate against base
Mixtral-8x7B-Instruct on AutoScientist's held-out marketing test set,
demonstrating meaningful improvement in marketing copy quality across
five task types.

## Environmental Impact

- **Hardware:** AutoScientist cloud infrastructure
- **Cloud Provider:** Adaption Labs
- **Training duration:** ~1 hour


## Technical Specifications

### Model Architecture and Objective

- **Base:** Mixtral-8x7B-Instruct-v0.1 (Mixture of Experts, 8 experts,
  2 active per token)
- **Adaptation:** LoRA adapters on all linear layers
- **Objective:** Supervised fine-tuning on marketing instruction data
- **Effective trainable parameters:** ~2% of base model parameters

### Compute Infrastructure

- **Platform:** AutoScientist by Adaption Labs
- **Method:** LoRA SFT with automatic recipe optimization

## Citation

If you use this model, please cite:

**BibTeX:**
```bibtex
@misc{huynh2026marketingmixtral,
  author = {Huynh, Nguyen},
  title = {Marketing-Mixtral-8x7B: A Fine-tuned Model for Marketing Copy Generation},
  year = {2026},
  publisher = {HuggingFace},
  howpublished = {\url{https://huggingface.co/suehuynh/Marketing-Mixtral-8x7B-v2}},
  note = {AutoScientist Challenge 2026 — Marketing Category}
}
```

## Model Card Authors

Sue Huynh — Brown University MSc Data Science

## Model Card Contact

HuggingFace: suehuynh