feat: Initial commit - Hermes Detective Agency concept
- Hermes Detective Agency: Open-ended mystery investigation game - Roles: Chief (human), Witness (Kimi), Detective (Hermes) - 5 difficulty levels, community cases, open-ended solving - Scoring: Alignment %, Evidence %, Time - Features: Retry, Journal, Observe mode - Tech: Kimi Vision + Hermes Agent + Pollinations Changelog: - Research phase: Kimi capabilities, Hermes agent, image APIs - Brainstorming: 14 ideas explored - Comparison matrix: Detective selected as winner - Concept finalized with all design decisions
This commit is contained in:
47
docs/research-hermes-agent.md
Normal file
47
docs/research-hermes-agent.md
Normal file
@@ -0,0 +1,47 @@
|
||||
# Research: Hermes Agent Capabilities
|
||||
|
||||
**Date:** 2026-04-19
|
||||
**Purpose:** Understand Hermes Agent framework for hackathon integration
|
||||
|
||||
## Hermes 3 (Nous Research)
|
||||
|
||||
### Core Capabilities
|
||||
- **Advanced agentic capabilities**
|
||||
- **Reliable function calling** - Trained specifically for tool use
|
||||
- **Structured output** - JSON mode / Pydantic schemas
|
||||
- **ChatML prompt format** - OpenAI-compatible
|
||||
- Multi-turn conversation
|
||||
- Long context coherence
|
||||
|
||||
### Benchmark Performance
|
||||
| Benchmark | Hermes 3 Score |
|
||||
|-----------|---------------|
|
||||
| IFEval (0-shot) | 61.70% |
|
||||
| MMLU-Redux | 92.7% |
|
||||
| MMLU-Pro | 81.1% |
|
||||
| SimpleQA | 31.0% |
|
||||
|
||||
### Function Calling
|
||||
- Trained on specific prompts for tool use
|
||||
- XML-based tool call format: `<tool_call>{"name": "...", "arguments": {...}}</tool_call>`
|
||||
- Supports recursive/chain tool calls
|
||||
- Native tool integration via NousResearch/Hermes-Function-Calling repo
|
||||
|
||||
## Hermes Agent Framework
|
||||
|
||||
### Key Components
|
||||
1. **ChatML format** - Structured system/user/assistant turns
|
||||
2. **Tool definitions** - JSON schema for function signatures
|
||||
3. **Tool parsing** - Parse and execute function calls
|
||||
4. **Response loop** - Multi-turn agentic execution
|
||||
|
||||
### Integration Points
|
||||
- HuggingFace Transformers
|
||||
- vLLM inference
|
||||
- Ollama local deployment
|
||||
- OpenAI-compatible API
|
||||
|
||||
## Sources
|
||||
- https://huggingface.co/NousResearch/Hermes-3-Llama-3.1-8B
|
||||
- https://github.com/NousResearch/Hermes-Function-Calling
|
||||
- https://arxiv.org/abs/2408.11857 (Hermes 3 Technical Report)
|
||||
Reference in New Issue
Block a user