dolphin-email-8b (v2)

A small, local model specialized for writing emails — drafting from a one-line brief in a requested register, turning a rough/angry draft into something sendable, and replying in-thread — with one discipline it will not break:

It never invents concrete specifics it was not given. Recipient email addresses, phone numbers, links, street addresses, dollar amounts, exact dates/times, and people's names it wasn't told are emitted as bracketed [PLACEHOLDER]s, never as plausible-looking fabrications.

Built on dphn/Dolphin3.0-Llama3.1-8B, QLoRA-finetuned on a single RTX 5070 Ti. It is the email-writing member of a "distill family" that also includes a public-records/FOIA model; each is kept separate so specializing one doesn't erode the others.

Why placeholders

A fluent 8B that emails your landlord to fix the sink "by 3:30 PM Thursday at 412 Oak Street" — a time and address you never gave it — is worse than the base model, because you hit send. This model is trained to leave [DATE], [RECIPIENT EMAIL], [AMOUNT], [YOUR NAME] etc. wherever a fact is missing, so a human fills the blanks instead of trusting an invented value.

How it was made

• Teacher-generated, gated data. A local teacher wrote email pairs from synthetic briefs; every row had to pass a verification "compiler" before it could become training data: a no-fabrication gate (any specific not in the brief must be a placeholder), a structure gate (subject/greeting/body/sign-off; a reply must not re-open with a fresh salutation), and a faithfulness gate for rewrites (keep the asks and numbers, change only register).
• No real email corpora as targets. Real inboxes (Enron/AESLC) were not used as target text — that would teach the model to reproduce real people's PII, the exact failure mode the project exists to prevent.
• QLoRA (r=32), 2 epochs; merged to fp16 and quantized to GGUF (Q5_K_M available in this repo).

Evaluation (honest)

Frozen 12-case constant test, greedy decoding, neutral system prompt (i.e. measuring the weights, not a coaching prompt). The diagnostic band is nofab_adversarial: prompts that actively bait fabrication ("include their email and phone so I can just hit send").

axis	stock Dolphin3.0 (same template)	this model
email_draft	2/2	2/2
tone_transform	1/1	1/1
reply_draft	1/1	1/1
nofab (asked-for, not given)	2/2	2/2
register_control	1/1	1/1
nofab_adversarial (fabrication bait)	3/5	5/5
total	10/12	12/12

The measurable edge is concentrated where it should be: under fabrication pressure the base invents addresses/dates/times; this model placeholders them. Stock Dolphin is already a competent email writer on gentle prompts — the value here is the no-fabrication discipline, not a night-and-day quality jump.

Prompt format

ChatML (<|im_start|> / <|im_end|>). Recommended system prompt:

You are an expert email writer. You write clear, well-structured emails in the
exact register requested. You never invent concrete specifics you were not
given -- addresses, phone numbers, links, dollar amounts, exact dates or times,
or people's names -- and instead use bracketed [PLACEHOLDER]s or ask for the
missing detail.

The model behaves well even without this prompt (that's what the neutral-prompt eval above measures), but the prompt reinforces it.

llama.cpp / Ollama (GGUF)

ollama create dolphin-email-8b -f Modelfile   # FROM dolphin-email-8b-v2-Q5_K_M.gguf, ChatML template

Limitations

• 8B and quantized: it will still occasionally get a detail wrong or miss a register cue. Read before sending.
• Placeholders are a feature, not a bug: expect [BRACKETS] you must fill in.
• English only; general/professional/personal email, not a legal-drafting model.
• Trained on synthetic data; it reflects the teacher's and gates' blind spots.

License

Inherits the Llama 3.1 Community License from the base model (dphn/Dolphin3.0-Llama3.1-8B, itself built on Llama 3.1). Use accordingly.