mjoshs/Qwen-Marketing-Turbine-Generic-Adapter

Turbine Qwen Marketing Adapters

Model summary

This repository contains a set of LoRA/PEFT adapters fine-tuned for B2B marketing and enablement copy in the workforce/apprenticeship domain, with optional persona specialization.

Adapters are trained on cleaned content extracted from a Turbine website codebase (MDX/MD and TSX pages), then shaped into persona-specific instruction-style examples.

Artifacts

• generic adapter: general Turbine marketing voice and product positioning.
• Persona adapters:
  • employers
  • sponsors
  • vocational_trainers
  • state_workforce_boards

These adapters are designed to be loaded onto the same base model:

• Base: marketeam/Qwen-Marketing

---

Intended use

Primary intended use

Generate, rewrite, and structure procurement-safe and persona-relevant marketing copy such as:

• landing page sections (headlines, bullets, CTAs)
• outbound emails and follow-ups
• one-pagers and enablement summaries
• objection handling
• benefit framing by stakeholder type

Out-of-scope use

Not intended for:

• legal advice or regulatory determinations
• guarantees of compliance, eligibility, or funding
• generating factual claims or metrics not present in the provided context
• sensitive personal data processing
• automated decision-making in high-stakes contexts (eligibility, enforcement)

---

Persona behaviors

Persona adapters are optimized for differences in framing and tone:

• Employers: ROI, speed-to-value, operational outcomes, reduced admin overhead.
• Sponsors: audit readiness, evidence capture, lifecycle operations, reporting artifacts.
• Vocational trainers: curriculum alignment, competency tracking, placements, employer alignment.
• State workforce boards: procurement-safe language, governance, interoperability, outcomes reporting, equity framing; avoids hype.

---

Training data

Source

Training inputs were derived from: https://turbineworkforce.com

• src/markdown/** (MDX/MD content)
• src/pages/** (TSX/TS content) with exclusions:
• paths containing blogPosts
• non-content markup and UI-only code

Cleaning and extraction

• HTML tags, MDX/JSX tags, and common TSX markup are removed.
• Frontmatter/imports/exports are stripped.
• Whitespace is normalized.
• Content is chunked into token windows for training.

Persona dataset generation

Persona datasets are auto-generated from non-persona text by applying persona-specific instruction frames. The resulting examples are chat-style records:

• system: persona writing constraints
• user: transformation task (rewrite / one-pager / email / objections)
• assistant: grounded content derived from the source text (no invented metrics)

Data files

Typical structure:

• out/splits/train.jsonl, out/splits/val.jsonl (non-persona)
• out/persona_splits/<persona>/train.jsonl, out/persona_splits/<persona>/val.jsonl

---

Training procedure

Method

• Parameter-efficient fine-tuning using LoRA (PEFT) on a causal LM.

Key hyperparameters (typical)

• LoRA:
  • target_modules: ["q_proj", "v_proj"]
  • r: 4–16 (persona-dependent)
  • lora_alpha: usually 2 * r
  • lora_dropout: 0.05–0.15 (higher for more conservative personas)
• Training:
  • max_length: ~512 tokens
  • effective_batch_size: ~16 (via gradient accumulation)
  • learning_rate: ~1e-4 (LoRA-appropriate)
  • mixed precision: fp16

Output

Only adapter weights are produced (base model weights are unchanged), unless explicitly merged.

---

Evaluation

What to evaluate

This is a marketing adapter set; success is measured by:

• persona separation: different stakeholders yield different framing
• groundedness: no invented metrics or unsupported claims
• style consistency: Turbine voice, clarity, procurement-safety where required
• format compliance: headline/bullets/CTA structure

Recommended evaluation

• Track eval_loss during training for overfit detection.
• Run fixed prompt suites per persona:
  • rewrite a paragraph → headline + bullets + CTA
  • produce an email + objection handling
  • procurement-safe one-pager section (boards)

---

Limitations

• Outputs may still contain generic marketing phrasing if persona data is small.
• The model can produce plausible-sounding claims if prompted to do so; users should provide source context and require citations/grounding in their pipeline if needed.
• Content quality depends on the source corpus quality; missing or thin pages lead to weaker examples.
• Not a compliance engine; it can mention regulations but should not be trusted to interpret them.

---

Safety and responsible use

• Do not use this model to fabricate metrics, outcomes, compliance guarantees, or legal statements.
• For state/board-facing outputs, enforce “procurement-safe” constraints and require that all claims be supported by provided context.
• Add a post-generation validation layer (rule-based or LLM-based) to flag:
  • numeric claims not present in context
  • absolute/guarantee language (“ensures”, “guaranteed”, “fully compliant”)
  • hallucinated program names or statutes

---

How to use

Load base + adapter (Transformers + PEFT)

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

BASE = "marketeam/Qwen-Marketing"
ADAPTER = "out/adapters/sponsors"  # or employers, vocational_trainers, state_workforce_boards

tok = AutoTokenizer.from_pretrained(BASE, trust_remote_code=True, use_fast=True)
model = AutoModelForCausalLM.from_pretrained(
    BASE, torch_dtype=torch.float16, device_map="auto", trust_remote_code=True
)

model = PeftModel.from_pretrained(model, ADAPTER)

prompt = "Write a sponsor-facing one-pager section about audit-ready evidence capture."
inputs = tok(prompt, return_tensors="pt").to(model.device)

out = model.generate(**inputs, max_new_tokens=300, do_sample=True, temperature=0.7)
print(tok.decode(out[0], skip_special_tokens=True))