praisonai-train 0.0.2__tar.gz

praisonai_train-0.0.2/PKG-INFO

@@ -0,0 +1,132 @@
+Metadata-Version: 2.4
+Name: praisonai-train
+Version: 0.0.2
+Summary: Training for PraisonAI — LLM fine-tuning (Unsloth) and iterative agent training extracted from the praisonai wrapper.
+Author: Mervin Praison
+License: MIT
+Project-URL: Homepage, https://docs.praison.ai
+Project-URL: Repository, https://github.com/mervinpraison/PraisonAI
+Requires-Python: <3.15,>=3.10
+Description-Content-Type: text/markdown
+Requires-Dist: praisonaiagents>=1.6.126
+Requires-Dist: rich>=13.7
+Requires-Dist: typer>=0.12.0
+Requires-Dist: click>=8.4.2
+Requires-Dist: PyYAML>=6.0
+Provides-Extra: llm
+Requires-Dist: praisonai-code>=0.0.1; extra == "llm"
+Requires-Dist: torch>=2.3.0; extra == "llm"
+Requires-Dist: transformers>=4.40.0; extra == "llm"
+Requires-Dist: unsloth>=2024.8; extra == "llm"
+Requires-Dist: trl<0.9.0; extra == "llm"
+Requires-Dist: peft>=0.10.0; extra == "llm"
+Requires-Dist: accelerate>=0.30.0; extra == "llm"
+Requires-Dist: bitsandbytes>=0.43.0; extra == "llm"
+Requires-Dist: datasets>=2.19.0; extra == "llm"
+Requires-Dist: sentencepiece>=0.2.0; extra == "llm"
+Requires-Dist: protobuf; extra == "llm"
+Requires-Dist: huggingface_hub>=0.23.0; extra == "llm"
+Requires-Dist: hf_transfer; extra == "llm"
+Provides-Extra: vision
+Requires-Dist: praisonai-train[llm]; extra == "vision"
+Provides-Extra: all
+Requires-Dist: praisonai-train[llm]; extra == "all"
+
+# PraisonAI Train
+
+Training for [PraisonAI](https://github.com/mervinpraison/PraisonAI) — fine-tune LLMs and iteratively train agents, as a standalone package or as part of the full `praisonai` stack.
+
+## What it does
+
+| Command | What happens | Needs GPU/ML deps? |
+|---------|--------------|--------------------|
+| `praisonai-train agents --input "What is Python?"` | Runs your agent, grades the answer with an LLM judge, feeds suggestions back, repeats | No |
+| `praisonai-train agents --input "Explain AI" --human` | Same loop, but **you** give the feedback | No |
+| `praisonai-train llm dataset.json` | Fine-tunes an open model (Llama, Qwen, …) on your dataset with [Unsloth](https://github.com/unslothai/unsloth) | Yes |
+| `praisonai-train list` / `show` / `apply` | Browse training sessions and apply the best iteration to an agent | No |
+
+## Install
+
+```bash
+# Agent training only (lightweight)
+pip install praisonai-train
+
+# + LLM fine-tuning (heavy ML stack: torch, unsloth, trl, ...)
+pip install "praisonai-train[llm]"
+
+# Or as part of the full PraisonAI stack (same commands via `praisonai train ...`)
+pip install "praisonai[train]"
+```
+
+GPU setups often prefer the conda installer, which pins CUDA-compatible versions:
+
+```bash
+setup-conda-env   # or: bash praisonai_train/setup/setup_conda_env.sh
+```
+
+## Quickstart: train an agent in 2 minutes
+
+```bash
+export OPENAI_API_KEY=sk-...
+
+# One input, three improvement iterations, LLM-as-judge
+praisonai-train agents --input "What is the capital of France?" --iterations 3
+
+# See what happened
+praisonai-train list
+praisonai-train show <session-id>
+
+# Apply the best iteration and chat with the improved agent
+praisonai-train apply <session-id> --run "And what about Germany?"
+```
+
+Python API:
+
+```python
+from praisonaiagents import Agent
+from praisonai_train import AgentTrainer, TrainingScenario
+
+agent = Agent(instructions="You are a helpful assistant.")
+trainer = AgentTrainer(agent=agent, iterations=3)
+trainer.add_scenario(TrainingScenario(id="demo", input_text="What is Python?"))
+report = trainer.run()
+report.print_summary()
+```
+
+## Quickstart: fine-tune an LLM
+
+```bash
+pip install "praisonai-train[llm]"
+
+# dataset.json in ShareGPT or Alpaca format; config.yaml is generated if absent
+praisonai-train llm dataset.json --model llama-3.1
+```
+
+Tuning knobs (LoRA rank, epochs, quantization, Ollama/HuggingFace export) live in `config.yaml` — see the template in `praisonai_train/setup/config.yaml`.
+
+## How it fits the PraisonAI stack
+
+```
+praisonaiagents  (core SDK)
+   ├── praisonai-code   (terminal CLI)
+   ├── praisonai-bot    (bots & gateway)
+   └── praisonai-train  (this package)
+        └── praisonai   (wrapper: installs everything)
+```
+
+- Depends only on `praisonaiagents` — no circular deps, installs standalone.
+- With the full stack installed, the same commands are available as `praisonai train ...`.
+- Old import paths (`praisonai.train.agents`, `python -m praisonai.train.llm.trainer`) keep working via wrapper shims.
+
+## Development
+
+```bash
+# From the monorepo root
+cd src/praisonai-train
+PYTHONPATH="../praisonai-agents:." python -m pytest tests/unit/train -q
+
+# Import-direction gate (train must not import the wrapper)
+bash ../../scripts/check_c10_train_imports.sh
+```
+
+Boundary details: `src/praisonai/tests/PRAISONAI_TRAIN_MANIFEST.md`.

praisonai_train-0.0.2/README.md

@@ -0,0 +1,98 @@
+# PraisonAI Train
+
+Training for [PraisonAI](https://github.com/mervinpraison/PraisonAI) — fine-tune LLMs and iteratively train agents, as a standalone package or as part of the full `praisonai` stack.
+
+## What it does
+
+| Command | What happens | Needs GPU/ML deps? |
+|---------|--------------|--------------------|
+| `praisonai-train agents --input "What is Python?"` | Runs your agent, grades the answer with an LLM judge, feeds suggestions back, repeats | No |
+| `praisonai-train agents --input "Explain AI" --human` | Same loop, but **you** give the feedback | No |
+| `praisonai-train llm dataset.json` | Fine-tunes an open model (Llama, Qwen, …) on your dataset with [Unsloth](https://github.com/unslothai/unsloth) | Yes |
+| `praisonai-train list` / `show` / `apply` | Browse training sessions and apply the best iteration to an agent | No |
+
+## Install
+
+```bash
+# Agent training only (lightweight)
+pip install praisonai-train
+
+# + LLM fine-tuning (heavy ML stack: torch, unsloth, trl, ...)
+pip install "praisonai-train[llm]"
+
+# Or as part of the full PraisonAI stack (same commands via `praisonai train ...`)
+pip install "praisonai[train]"
+```
+
+GPU setups often prefer the conda installer, which pins CUDA-compatible versions:
+
+```bash
+setup-conda-env   # or: bash praisonai_train/setup/setup_conda_env.sh
+```
+
+## Quickstart: train an agent in 2 minutes
+
+```bash
+export OPENAI_API_KEY=sk-...
+
+# One input, three improvement iterations, LLM-as-judge
+praisonai-train agents --input "What is the capital of France?" --iterations 3
+
+# See what happened
+praisonai-train list
+praisonai-train show <session-id>
+
+# Apply the best iteration and chat with the improved agent
+praisonai-train apply <session-id> --run "And what about Germany?"
+```
+
+Python API:
+
+```python
+from praisonaiagents import Agent
+from praisonai_train import AgentTrainer, TrainingScenario
+
+agent = Agent(instructions="You are a helpful assistant.")
+trainer = AgentTrainer(agent=agent, iterations=3)
+trainer.add_scenario(TrainingScenario(id="demo", input_text="What is Python?"))
+report = trainer.run()
+report.print_summary()
+```
+
+## Quickstart: fine-tune an LLM
+
+```bash
+pip install "praisonai-train[llm]"
+
+# dataset.json in ShareGPT or Alpaca format; config.yaml is generated if absent
+praisonai-train llm dataset.json --model llama-3.1
+```
+
+Tuning knobs (LoRA rank, epochs, quantization, Ollama/HuggingFace export) live in `config.yaml` — see the template in `praisonai_train/setup/config.yaml`.
+
+## How it fits the PraisonAI stack
+
+```
+praisonaiagents  (core SDK)
+   ├── praisonai-code   (terminal CLI)
+   ├── praisonai-bot    (bots & gateway)
+   └── praisonai-train  (this package)
+        └── praisonai   (wrapper: installs everything)
+```
+
+- Depends only on `praisonaiagents` — no circular deps, installs standalone.
+- With the full stack installed, the same commands are available as `praisonai train ...`.
+- Old import paths (`praisonai.train.agents`, `python -m praisonai.train.llm.trainer`) keep working via wrapper shims.
+
+## Development
+
+```bash
+# From the monorepo root
+cd src/praisonai-train
+PYTHONPATH="../praisonai-agents:." python -m pytest tests/unit/train -q
+
+# Import-direction gate (train must not import the wrapper)
+bash ../../scripts/check_c10_train_imports.sh
+```
+
+Boundary details: `src/praisonai/tests/PRAISONAI_TRAIN_MANIFEST.md`.

praisonai_train-0.0.2/praisonai_train/__init__.py

@@ -0,0 +1,39 @@
+"""PraisonAI Train — LLM fine-tuning and agent training (Tier 2c).
+
+Extracted from the ``praisonai`` wrapper (C10). Two training modes:
+
+- **LLM fine-tuning** (``praisonai-train llm``): Unsloth-based fine-tuning of
+  open models, driven by a ``config.yaml``. Heavy ML deps are lazy — install
+  with ``pip install "praisonai-train[llm]"`` or the conda script in
+  ``praisonai_train/setup/``.
+- **Agent training** (``praisonai-train agents``): iterative LLM-as-judge or
+  human-in-the-loop feedback loops. Needs only ``praisonaiagents``.
+
+Old wrapper import paths (``praisonai.train.*``) keep working via shims.
+"""
+
+from __future__ import annotations
+
+from praisonai_train._version import __version__
+
+__all__ = [
+    "__version__",
+    "AgentTrainer",
+    "TrainingScenario",
+    "TrainModel",
+    "apply_training",
+    "get_training_profile",
+]
+
+
+def __getattr__(name: str):
+    """Lazy top-level exports — never import ML deps at package import time."""
+    if name in ("AgentTrainer", "TrainingScenario", "apply_training", "get_training_profile"):
+        from praisonai_train.train import agents as _agents
+
+        return getattr(_agents, name)
+    if name == "TrainModel":
+        from praisonai_train.train.llm import trainer as _trainer
+
+        return _trainer.TrainModel
+    raise AttributeError(f"module {__name__!r} has no attribute {name!r}")

praisonai_train-0.0.2/praisonai_train/__main__.py

@@ -0,0 +1,18 @@
+"""Console entry: ``praisonai-train``."""
+
+from __future__ import annotations
+
+import sys
+
+
+def main(argv: list[str] | None = None) -> None:
+    from praisonai_train.cli.app import app
+
+    args = argv if argv is not None else sys.argv[1:]
+    if not args:
+        args = ["--help"]
+    app(args=args, prog_name="praisonai-train")
+
+
+if __name__ == "__main__":
+    main()

praisonai_train-0.0.2/praisonai_train/_bootstrap.py

@@ -0,0 +1,37 @@
+"""Monorepo bootstrap for ``praisonai_train`` and optional ``praisonai_code``."""
+
+from __future__ import annotations
+
+import sys
+from pathlib import Path
+
+
+def ensure_praisonai_train() -> None:
+    """Ensure ``praisonai_train`` is importable in monorepo dev layouts."""
+    try:
+        import praisonai_train  # noqa: F401
+        return
+    except ImportError:
+        pass
+
+    here = Path(__file__).resolve().parents[1]  # .../src/praisonai-train
+    if (here / "praisonai_train").is_dir():
+        root = str(here)
+        if root not in sys.path:
+            sys.path.insert(0, root)
+
+
+def ensure_praisonai_code() -> None:
+    """Optional code-tier imports (YAML agent loading, legacy dispatch) when co-installed."""
+    try:
+        import praisonai_code  # noqa: F401
+        return
+    except ImportError:
+        pass
+
+    train_src = Path(__file__).resolve().parents[1]
+    code_src = train_src.parent / "praisonai-code"
+    if (code_src / "praisonai_code").is_dir():
+        root = str(code_src)
+        if root not in sys.path:
+            sys.path.insert(0, root)

praisonai_train-0.0.2/praisonai_train/_code_bridge.py

@@ -0,0 +1,28 @@
+"""Lazy access from praisonai-train to optional praisonai-code modules."""
+
+from __future__ import annotations
+
+import importlib
+from typing import Any
+
+from praisonai_train._bootstrap import ensure_praisonai_code
+
+
+def code_available() -> bool:
+    ensure_praisonai_code()
+    try:
+        import praisonai_code  # noqa: F401
+        return True
+    except ImportError:
+        return False
+
+
+def import_code_module(name: str) -> Any:
+    ensure_praisonai_code()
+    try:
+        return importlib.import_module(name)
+    except ImportError as exc:
+        raise ImportError(
+            f"Optional code module {name!r} requires praisonai-code. "
+            "Install with: pip install praisonai-code"
+        ) from exc

praisonai_train-0.0.2/praisonai_train/_version.py

@@ -0,0 +1 @@
+__version__ = "0.0.2"

praisonai_train-0.0.2/praisonai_train/_wrapper_bridge.py

@@ -0,0 +1,31 @@
+"""Lazy access from praisonai-train to optional praisonai wrapper modules."""
+
+from __future__ import annotations
+
+import importlib
+from typing import Any
+
+
+def wrapper_available() -> bool:
+    try:
+        import praisonai  # noqa: F401
+        return True
+    except ImportError:
+        return False
+
+
+def import_wrapper_module(name: str) -> Any:
+    if not wrapper_available():
+        raise ImportError(
+            f"Optional wrapper module {name!r} requires the praisonai package. "
+            "Install with: pip install praisonai"
+        )
+    return importlib.import_module(name)
+
+
+def optional_wrapper_attr(module: str, attr: str, default: Any = None) -> Any:
+    try:
+        mod = import_wrapper_module(module)
+        return getattr(mod, attr)
+    except ImportError:
+        return default

praisonai_train-0.0.2/praisonai_train/cli/__init__.py

@@ -0,0 +1 @@
+"""CLI for standalone ``praisonai-train``."""

praisonai_train-0.0.2/praisonai_train/cli/app.py

@@ -0,0 +1,15 @@
+"""Typer CLI for standalone ``praisonai-train``.
+
+The train command group (``llm``, ``agents``, ``list``, ``show``, ``apply``)
+is a single Typer app, so the standalone console script exposes it directly:
+``praisonai-train llm dataset.json``, ``praisonai-train agents --input ...``.
+
+Inside the full stack the same app is mounted as ``praisonai train`` via the
+``_TRAIN_RESIDENT_COMMANDS`` routing in ``praisonai_code.cli.app``.
+"""
+
+from __future__ import annotations
+
+from praisonai_train.cli.commands.train import app
+
+__all__ = ["app"]

praisonai_train-0.0.2/praisonai_train/cli/commands/__init__.py

@@ -0,0 +1 @@
+"""Typer command modules for praisonai-train."""