h0ney-badger/dolphin-email-8b

TEXT GENERATIONPricing:Input $0.37 / Cached $0.074 / Output $0.38Concurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:8kTool Calling:SupportedPublished:Sep 8, 2026License:llama3.1Architecture:Transformer Featherless Exclusive Cold

h0ney-badger/dolphin-email-8b is an 8 billion parameter language model, QLoRA-finetuned on dphn/Dolphin3.0-Llama3.1-8B, specifically designed for email writing tasks. Its key characteristic is a strict no-fabrication policy, replacing unprovided concrete specifics (like addresses, dates, or names) with bracketed placeholders to prevent the invention of false information. This model excels at drafting emails from briefs, transforming rough drafts, and replying in-thread, ensuring human review for critical details.

Loading preview...

Overview

h0ney-badger/dolphin-email-8b is an 8 billion parameter model, QLoRA-finetuned from dphn/Dolphin3.0-Llama3.1-8B, specialized in email writing. Its core differentiator is a strict no-fabrication policy: it will never invent concrete specifics (e.g., recipient email addresses, phone numbers, links, dollar amounts, exact dates/times, or people's names) that were not explicitly provided. Instead, it inserts bracketed [PLACEHOLDER]s, ensuring human intervention for critical details.

Key Capabilities

  • Email Drafting: Generates emails from one-line briefs in a requested register.
  • Tone Transformation: Rewrites rough or angry drafts into professional, sendable messages.
  • In-thread Replies: Drafts replies that maintain conversational flow.
  • No Fabrication: Replaces unprovided concrete details with [PLACEHOLDER]s, preventing the generation of false information, especially under adversarial prompts.

How it was Made

The model was trained using QLoRA on teacher-generated, gated data. This data underwent a rigorous verification process to ensure no fabrication, correct email structure, and faithfulness for rewrites. Crucially, no real email corpora were used as target text to avoid reproducing personally identifiable information.

Evaluation Highlights

In a 12-case constant test with greedy decoding and a neutral system prompt, this model achieved a perfect 12/12 score, outperforming the stock Dolphin3.0 (10/12). The significant improvement was observed in nofab_adversarial cases, where the base model invented details while dolphin-email-8b correctly used placeholders.

Good for

  • Users who need a reliable AI assistant for drafting and refining emails.
  • Applications requiring email generation where factual accuracy and the prevention of invented details are paramount.
  • Scenarios where a human must review and fill in specific, sensitive information before sending.

Limitations

  • Occasional errors in detail or register are possible; human review is always recommended.
  • Expect [BRACKETS] for missing information, as this is an intentional feature.
  • English only, focused on general/professional/personal email, not legal drafting.
  • Trained on synthetic data, reflecting the biases of its creation process.

Finetunes
1 models