RYVR/qwen3-32b-b2b-marketing-lora

Qwen3-32B B2B Marketing (LoRA) — v2

A LoRA adapter for Qwen/Qwen3-32B, fine-tuned for brief-driven B2B marketing copywriting: ABM and cold emails, LinkedIn posts, blog intros, Google ads, landing-page heroes, nurture sequences, event invites, follow-ups, and case-study summaries.

Successor to RYVR/qwen2.5-7b-b2b-marketing-lora. Developed by RYVR.

What's different in v2

• Brief-conditioned: trained on structured briefs (company, ICP, pain points, tone, hard constraints, CTA) rather than bare instructions — give it a real brief and it follows the constraints.
• Anti-fabrication training: 60% of training briefs explicitly forbid invented statistics; the model is trained to use only numbers supplied in the brief.
• Quality-gated data: every training pair was scored by an LLM judge; only high-scoring pairs (median 9/10) were kept.

Evaluation

Blind LLM-judged benchmark (50 prompts: 25 short-instruction + 25 real-world briefs; judges never saw model identities; both models generated on identical hardware):

Metric	Qwen3-32B base	This adapter
Overall score (0–10)	4.79	7.16
Head-to-head win rate vs base	—	86% (43W/6L/1T)
Brief-constraint adherence	6.16	7.76
Fabricated statistics (per 50 outputs)	13	4
Truncations / format breaks	12	0

How to use

Recommended: run with thinking disabled (the adapter was trained in non-thinking mode).

🤗 PEFT + Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_id = "Qwen/Qwen3-32B"
adapter_id = "RYVR/qwen3-32b-b2b-marketing-lora"

tokenizer = AutoTokenizer.from_pretrained(adapter_id)
model = AutoModelForCausalLM.from_pretrained(base_id, torch_dtype="auto", device_map="auto")
model = PeftModel.from_pretrained(model, adapter_id)

brief = """Write a personalised ABM email using this brief.

Company: Northwind Logistics, a freight-visibility platform
ICP: Head of Supply Chain at mid-market importers
Pain points: blind spots between ports; demurrage fees
Tone: direct, operator-to-operator
Constraints: under 130 words; one clear ask
CTA: a 15-minute lane-visibility audit
Usable facts: none — no invented statistics."""

messages = [
    {"role": "system", "content": "You are an expert B2B marketing copywriter."},
    {"role": "user", "content": brief},
]
inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True,
                                       enable_thinking=False, return_tensors="pt").to(model.device)
outputs = model.generate(inputs, max_new_tokens=700, temperature=0.7, top_p=0.9)
print(tokenizer.decode(outputs[0][inputs.shape[1]:], skip_special_tokens=True))

vLLM (LoRA runtime)

vllm serve Qwen/Qwen3-32B \
  --enable-lora \
  --lora-modules b2b-marketing=RYVR/qwen3-32b-b2b-marketing-lora

The brief format

The adapter performs best with structured briefs. Fields it was trained on: Company, Product, ICP, Pain points, Tone, Constraints (word limits, structure, banned words, required mentions), CTA, and Usable facts (the only numbers the copy may cite — write "none" to get statistics-free copy).

Training details

• Base model: Qwen/Qwen3-32B (non-thinking mode)
• Method: LoRA SFT — r=16, alpha=32, target modules q/k/v/o/gate/up/down_proj
• Data: 7,722 brief→asset pairs across 10 asset types, 48 industries, ~20 personas; multi-teacher generated, LLM-judge filtered (score ≥7/10), deduplicated; final training loss ≈ 1.03 over 2 epochs (5.1M tokens)
• Training data is proprietary and not distributed with this adapter

Intended use & limitations

For drafting B2B marketing copy from briefs. Human review is required before publishing: the model is trained against inventing statistics, but no model is immune — verify any factual or numeric claim. Predominantly English, western business idiom. Not for legal, medical, or financial advice.

License

Apache 2.0, consistent with the Qwen3-32B base model.