minds-hamlet 0.1.0

Lightweight all-in-one framework for building, instrumenting, and fine-tuning LLM code agents.

HAMLET is a lightweight, all-in-one framework for building and iterating on LLM agents. It lets you define agents, launch a Gradio GUI, instrument runs with Langfuse, and fine-tune models using the built-in GRPO trainer. The framework is modified primarily on two open-source projects: smolagents for the core agent structure (src/hamlet/core) and verifiers for the training stack (src/hamlet/train).

Installation

Install the published package when you just need the library, or clone the repo if you plan to modify HAMLET.

Option A: Install the package

# inside your project
uv add minds-hamlet

# or with pip
pip install minds-hamlet

This exposes the hamlet Python package for imports.

Option B: Work from source

Follow the steps below to set up a local development environment. The project uses uv to manage Python dependencies because it keeps lock files fast and reproducible. You can still use pip, but uv is the recommended path.

1. Prerequisites

• Python 3.10 or newer (3.11+ works as well)
• uv (install with the command below or follow the instructions in the uv repository)

pip install uv

2. Clone the repository

git clone https://github.com/MINDS-THU/HAMLET.git
cd HAMLET

3. Install dependencies

Install the base runtime:

uv sync

Optional extras:

• uv sync --extra tools for the toolchain utilities (file editing, retrieval, visual QA, etc.).
• uv sync --extra train for the training stack (GRPO trainer, vLLM client, etc.).

4. Use the environment

• Run commands inside the uv-managed env with uv run, e.g. uv run pytest or uv run python examples\gradio_gui_example.py.
• Alternatively, activate the virtual environment directly: .\.venv\Scripts\activate on Windows or source .venv/bin/activate on Unix shells.

5. Configure API keys

Create a .env file (or export environment variables) with the credentials your agent or tools need. Common entries include OPENAI_API_KEY, OPENAI_BASE_URL, and any service-specific tokens (HuggingFace, Langfuse, etc.). The examples load this file via dotenv, so keep it at the repo root and never commit it.

Getting Started

All commands below assume you are in the repository root after running uv sync.

1. Gradio GUI walkthrough

uv run python examples\gradio_gui_example.py

Launches the interactive UI so you can chat with an agent, observe tool traces, and any artifact generated by the agent.

2. Parallel code blocks

uv run python examples\parallel_code_blocks_example.py

Shows how HAMLET dispatches multiple LLM-generated code blocks concurrently and uses an optional Early Stop Strategy: code snippet to stop remaining executions once a shared goal is met.

Example run:

Thought:
I'll compute the sum of squares 1^2 + 2^2 + ... + 75^2 using two independent strategies:
1) Closed-form formula: n(n+1)(2n+1)/6.
2) Direct summation (Python sum over a generator).
I'll print the computed totals (using the same variable name total in both blocks). I'll use Early Stop Strategy: code to check that the computed total lies within the 0–500000 bound (0 <= total <= 500000). 

Code#1:
<code>
# Code#1: closed-form formula for sum of squares
n = 75
total = n * (n + 1) * (2 * n + 1) // 6
print("Method: formula, total:", total)
</code>

Code#2:
<code>
# Code#2: direct iterative summation
total = sum(i * i for i in range(1, 76))
print("Method: iterative sum, total:", total)
</code>

Early Stop Strategy: code

Early Stop Code:
<code>
# Early Stop Code: check whether 'total' is within the requested 0-500000 bound
early_stop_result = (0 <= total <= 500000)
</code>

Log excerpt:

Early stop strategy detected: 'code'. Will check each code execution result accordingly.
Execution logs for Code#1 (2 in total):
Method: formula, total: 143450
Executing parsed early stop code for Code#1 (2 in total)...
Evaluation for the execution result of Code#1 passed (2 in total), other code executions will be cancelled.

3. Structured schema I/O

uv run python examples\structured_schema_example.py

Demonstrates attaching JSON schemas to agent inputs and outputs so downstream automation receives validated, typed payloads. The sample defines:

class PersonaRequest(BaseModel):
	persona: str
	goal: str
	tone: str

class PersonaResponse(BaseModel):
	summary: str
	talking_points: list[str]
	next_steps: list[str]

HAMLET hands these schemas to the LLM, validates the JSON that comes back, and surfaces parsing errors if the reply drifts from the contract.

Tip: prepend uv run to each script while uv manages the environment, or activate .venv if you prefer calling python directly.