关于使用OpenAI API生成合规TL/DR摘要的技术咨询

Hey Mike, I’ve run into similar constraints with custom summary generation using OpenAI’s APIs, so let’s walk through how to get your TL/DR outputs exactly right—starting with prompt tweaks (no need for fine-tuning immediately if you nail this).

1. Super-Specific Prompt Constraints (The First Fix) #

The key here is to leave zero ambiguity for the model. You need to spell out every requirement explicitly, including how to handle edge cases like bullet points and short texts. Here’s a tested prompt template you can adapt:

Generate a strict TL/DR summary of the text below that follows these rules exactly:
- Must be 1 OR 2 complete, grammatically correct sentences (no fragments)
- Total character count (including spaces and punctuation) must NOT exceed 250
- If the original text has bullet points, rephrase all list content into flowing, natural sentences—do NOT copy or reference bullet points directly
- If the original text is short, use only 1 sentence unless a second sentence adds unique, non-redundant context (avoid repeating the same idea with minor wording changes)
---
[Paste your original text here]

Pair this with parameter adjustments:

• Set temperature to 0.2–0.5: Lower values reduce randomness, which helps avoid redundant sentences and off-format outputs.
• Set max_tokens to 150–200: Since 1 token ≈ 4 characters, this gives the model enough space to finish sentences without exceeding your 250-character limit (the prompt’s character rule will keep it in check).
• Add presence_penalty: 0.1–0.3: This discourages the model from getting stuck on partial sentences or repetitive phrasing, which helps with truncation issues.

2. Targeted Fixes for Your Specific Pain Points #

Truncated Sentences #

Stop relying solely on max_tokens to control length—your prompt’s explicit “complete sentences” rule is more effective. The max_tokens value above gives the model breathing room to finish sentences, while the character limit ensures it doesn’t ramble. If you still see truncation, add a line to the prompt: "Do NOT cut off sentences mid-way; ensure every sentence ends properly with punctuation."

Bullet Point Extraction #

Add a concrete example to your prompt to show exactly how you want lists converted. For instance, insert this right after the rules:

Example: Original bullet points: "- Resolve payment gateway errors - Add multi-language support - Update privacy policy" → Summary: "The update addresses payment gateway errors, introduces multi-language support, and includes revisions to the privacy policy."

This gives the model a clear blueprint for rephrasing lists into sentences.

Redundant Short-Text Summaries #

Double down on the prompt’s rule about avoiding repetition, and set temperature to the lower end (0.2–0.3). This reduces the model’s urge to “pad” short texts with minor variations of the same idea. If the text is truly too short for two sentences, the model will default to one—exactly what you want.

3. When to Consider Fine-Tuning #

If you’ve tested multiple prompt variations and parameter tweaks and still aren’t getting consistent results, fine-tuning with 1000+ high-quality examples is absolutely a solid plan.

1000+ examples is more than enough for the model to learn your exact formatting and content rules—OpenAI recommends starting with at least a few hundred examples, so 1000+ will give you strong, consistent outputs. Make sure each training example follows this structure (save as JSONL):

{"prompt": "[Original text here]\n\nTL/DR:", "completion": " [Your perfect 1-2 sentence, <250 character summary here>"}

The \n\nTL/DR: signal helps the model understand where the prompt ends and the summary task begins.

Final Takeaway #

Start with the hyper-specific prompt and parameter adjustments—this is the fastest, cheapest way to fix most of your issues. If you still need tighter consistency, your 1000+ example dataset will work great for fine-tuning. You’ll be getting exactly the TL/DR summaries you want in no time!

内容的提问来源于stack exchange，提问作者Mike Smith