--- license: gemma base_model: google/gemma-4-31b-it library_name: transformers pipeline_tag: text-generation language: - en tags: - copywriting - marketing - advertising - ad-copy - sales-copy - email-marketing - direct-response - conversion-copywriting - gemma - gemma-4 - text-generation - emotional-intelligence - conversational --- # Copywriter-Gemma4-31B **An AI copywriting LLM model that writes like a senior direct-response copywriter — not a chatbot.** Copywriter-Gemma4-31B is a 31-billion-parameter large language model fine-tuned from **Google Gemma-4-31B-it** for one job done exceptionally well: **writing high-converting marketing copy**. Facebook and Google ads, cold emails, landing pages, product descriptions, sales pages, video scripts, taglines, SMS campaigns, and more — in a punchy, specific, human voice that sells. In a blind, order-balanced head-to-head judged by **DeepSeek V4 Flash**, it **wins 80% of matchups against its own base model** and lifts its preference Elo by **+290 points**. --- ## TL;DR | | | |---|---| | 🧠 **What it is** | A Gemma-4-31B model specialized for copywriting & marketing | | 🏆 **Headline result** | **Elo 1657 vs 1367** for base — **wins 24 of 30** head-to-head briefs (80%) | | ⚖️ **Judge** | DeepSeek V4 Flash, blind pairwise (both orderings) | | ✍️ **Best at** | Ad copy, email, landing pages, hooks, product copy, scripts | | 📐 **Context** | 256K tokens · bfloat16 · drops into vLLM / Transformers | --- ## Why this model exists General-purpose chat models *can* write copy, but they default to the same tells: hedging, throat-clearing intros, vague benefit-speak, and "In today's fast-paced world..." openers. Copywriter-Gemma4-31B was trained to do what great copywriters do instead — **lead with the pain, get specific, cut the fluff, and earn the click.** It was fine-tuned on a large, curated corpus of marketing copy that includes **thousands of gold-standard, real-world ad examples**, so it has internalized the patterns of copy that actually performs rather than copy that merely sounds polished. --- ## Benchmark Results Most model cards make claims. This one shows the scoreboard. We built a copywriting-specific evaluation on top of the **EQ-Bench 3** methodology (pairwise Elo + a multi-dimension rubric), swapping in **30 real-world copywriting briefs** spanning Facebook ads, cold email, LinkedIn posts, landing pages, product descriptions, SMS, scripts, press releases, and more. Every brief was answered by both the **base** model and the **fine-tuned** model, then scored by an independent **DeepSeek V4 Flash** judge. To remove position bias, **every matchup was judged in both orderings** (A-vs-B and B-vs-A). Both models ran on the *identical* base weights and decoding settings — the only variable is the fine-tune. ### Main metric — preference Elo | Model | Elo | Head-to-head vs base | |---|---:|---| | **Copywriter-Gemma4-31B** | **1657** 🥇 | **wins 24 / 30 (80%)** | | Gemma-4-31B-it (base) | 1367 | — | **+290 Elo.** When a neutral judge is asked *"which piece of copy is better?"*, it picks Copywriter-Gemma4-31B **four times out of five.** ### Ability profile Across the rubric's seven copywriting dimensions, the fine-tune shows its biggest, most consistent gains exactly where direct-response copy lives: - ✅ **Hook strength** — stronger, scroll-stopping openers - ✅ **Specificity** — concrete details over generic benefit-speak - ✅ **Concision** — tighter copy, less filler > **Method notes:** judge = `deepseek-v4-flash`; 30 briefs; pairwise scoring in both orderings plus per-dimension rubric. Both the base and the fine-tune were served **identically** — same 31B base, same 4-bit quantization at inference, same decoding — so the only variable is the fine-tune. The weights published here are that **same fine-tune merged to full bf16 precision** (the higher-fidelity form of the model that produced these scores). Elo is computed *relative to the base model* on this brief set — a controlled A/B signal, not a public-leaderboard number. ### ⚡ Achieved in direct mode — no "thinking" required - It's Bad Every result above was generated in **direct (non-thinking) mode** — `enable_thinking=false` (the default), so the model wrote finished copy in a **single pass with zero reasoning tokens.** No chain-of-thought overhead, no extra latency or cost — it just writes. Gemma 4 *does* support an optional reasoning mode (`enable_thinking=True` → ` ` channel) if you want it for unusually complex briefs, but the scores above prove it isn't needed for strong copy. Fast by default. NOTE: Thinking must be off for best results. --- ## What it's good at Copywriter-Gemma4-31B is built for **marketing and growth teams, agencies, founders, and solo creators** who need on-brand copy fast: - **Paid ads** — Facebook / Instagram / Meta ad copy, Google responsive search ads, primary text + headlines - **Email** — cold outreach, welcome sequences, re-engagement, abandoned-cart, newsletters - **Web** — landing pages, hero sections, value props, feature/benefit bullets, CTAs - **Ecommerce** — product titles & descriptions, Amazon bullets, offer stacks - **Social** — hooks, LinkedIn posts, short-form video scripts, captions, taglines - **Lifecycle** — SMS campaigns, push notifications, upsell/order-bump copy --- ## Model details | | | |---|---| | **Base model** | [google/gemma-4-31b-it](https://huggingface.co/google/gemma-4-31b-it) | | **Parameters** | ~31B | | **Architecture** | Gemma 4 (`Gemma4ForConditionalGeneration`) | | **Context length** | 262,144 tokens (256K) | | **Precision** | bfloat16 | | **Format** | Safetensors (sharded) — vLLM / Transformers ready | | **Fine-tuning** | Supervised fine-tuning (QLoRA) on a curated copywriting corpus | | **Language** | English | | **License** | Gemma (see below) | --- ## Training Copywriter-Gemma4-31B was produced by **supervised fine-tuning** of Gemma-4-31B-it on a large, purpose-built copywriting dataset assembled and cleaned specifically for this project. The corpus pairs marketing briefs with high-quality completions and includes **thousands of gold-standard, real-world advertisements** so the model learns the rhythm, structure, and restraint of copy that converts — strong hooks, concrete specifics, tight CTAs, and a human voice. The final weights are the **full fine-tuned model merged to bfloat16** (not an adapter), so it loads like any standard Hugging Face model. --- ## Usage ### vLLM (recommended for serving) ```bash vllm serve akwin123/copywriter-gemma4-31b --dtype bfloat16 ``` ### Transformers ```python import torch from transformers import AutoModelForCausalLM, AutoTokenizer model_id = "akwin123/copywriter-gemma4-31b" tokenizer = AutoTokenizer.from_pretrained(model_id) model = AutoModelForCausalLM.from_pretrained( model_id, torch_dtype=torch.bfloat16, device_map="auto", ) brief = ( "Write a punchy Facebook ad for a budgeting app aimed at college students. " "Lead with the pain point and finish with a clear call to action. Keep it tight." ) messages = [{"role": "user", "content": brief}] inputs = tokenizer.apply_chat_template( messages, return_tensors="pt", add_generation_prompt=True ).to(model.device) outputs = model.generate( inputs, max_new_tokens=512, do_sample=True, temperature=0.7, min_p=0.1 ) print(tokenizer.decode(outputs[0][inputs.shape[-1]:], skip_special_tokens=True)) ``` > *Note: Gemma 4 is a multimodal architecture. If `AutoModelForCausalLM` does not resolve on your Transformers version, load it via the `Gemma4ForConditionalGeneration` class — text generation works the same way.* ### Recommended decoding | Setting | Value | Why | |---|---|---| | `temperature` | `0.7` | lively but on-brief | | `min_p` | `0.1` | trims low-probability noise | | `max_new_tokens` | `256–1024` | by format | --- ## Prompting tips - **Give it the brief, not just the product.** Audience, offer, channel, tone, and the *one* action you want — the more context, the sharper the copy. - **Ask for the angle.** "Lead with the pain," "make it curiosity-driven," "benefit-first," "contrarian hook." - **Ask for variations.** It's strong at generating multiple distinct headlines/subject lines you can A/B test. - **Specify length and format** (e.g., "3 SMS messages under 160 characters," "a 30-second VO script"). --- ## Limitations & scope - **English, marketing-focused.** Outside copywriting/marketing it behaves like a general Gemma-4 model. - **Occasional repetition** on long enumerations (e.g., "give me 10 taglines"); a touch of `temperature`/`min_p` or asking for fewer, better options resolves it. - **Not a fact-checker.** It will write whatever claims the brief implies — review for accuracy, legal/regulatory compliance, and brand safety before publishing. Always keep a human in the loop. - **Not for medical, legal, financial, or safety-critical advice.** --- ## License This model is derived from Google's Gemma-4 and is released under the **[Gemma Terms of Use](https://ai.google.dev/gemma/terms)**. By using these weights you agree to the Gemma license and Google's [Prohibited Use Policy](https://ai.google.dev/gemma/prohibited_use_policy). Use must comply with all upstream terms. --- ## Dataset licensing & commercial inquiries The **curated copywriting dataset** behind this model — instruction and preference data and more, including **thousands of gold-standard real-world ad examples** — is **available for licensing**. It's the same proven corpus that produced the benchmark results above, ideal for teams training their own marketing/copywriting models. 📩 **For dataset licensing, custom fine-tunes, or commercial partnerships, get in touch: [akash@adlavishmedia.com](mailto:akash@adlavishmedia.com)** --- ## About Copywriter-Gemma4-31B is an independent fine-tune focused on practical, high-performing marketing copy. If you build with it, benchmark it, or find failure cases, contributions and feedback are welcome on the repository's Community tab. *Built on Gemma-4-31B-it · evaluated with an EQ-Bench-3-style pairwise harness · judged by DeepSeek V4 Flash.*
