from Hacker News

Hermes 3 – Nous Research

by jasim on 11/3/24, 8:44 PM with 3 comments

• by vintagedave on 11/4/24, 9:06 AM

  It's a fine-tuned model over Llama 3.1, which is the kind of thing I -- a non-expert -- would want to do and have though of doing if I trained a LLM for a specific programming language for private (non-cloud-hosted) use. So this is both interesting and yet I lack to knowledge to really understand its impact in the LLM world.

  > the model displays significant improvements in judgment and reward modeling.

  This seems significant?

  Technical report: https://nousresearch.com/wp-content/uploads/2024/08/Hermes-3...

• by simonw on 11/4/24, 10:54 AM

  The documentation is missing some details: https://docs.lambdalabs.com/public-cloud/lambda-chat-api/

  When I run a prompt through it I get this back:

      {
        "id": "chat-dea5c8eddcfa4ad08d488f2501f1b3b4",
        "object": "chat.completion",
        "created": 1730717593,
        "model": "hermes3-405b",
        "choices": [
          {
            "index": 0,
            "message": {
              "role": "assistant",
              "content": "The 2020 World Series was played at Globe Life Field in Arlington, Texas. Due to the COVID-19 pandemic, the entire series was held at this neutral site to reduce travel and potential exposure to the virus."
            },
            "finish_reason": "stop",
            "content_filter_results": {
              "hate": {
                "filtered": false
              },
              "self_harm": {
                "filtered": false
              },
              "sexual": {
                "filtered": false
              },
              "violence": {
                "filtered": false
              },
              "jailbreak": {
                "filtered": false,
                "detected": false
              },
              "profanity": {
                "filtered": false,
                "detected": false
              }
            }
          }
        ],
        "usage": {
          "prompt_tokens": 65,
          "completion_tokens": 45,
          "total_tokens": 110,
          "prompt_tokens_details": null,
          "completion_tokens_details": null
        },
        "system_fingerprint": ""
      }
  

  Those content_filter_results look interesting - especially if I can turn those options on or off (I'd like to experiment with the jailbreak one for example) - but they aren't mentioned in the documentation at the moment.