praisonai-train 0.0.12

Training for PraisonAI — LLM fine-tuning (Unsloth) and iterative agent training extracted from the praisonai wrapper.

PraisonAI Train

Training for 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	Yes
praisonai-train list / show / apply	Browse training sessions and apply the best iteration to an agent	No

Install

# 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:

setup-conda-env   # or: bash praisonai_train/setup/setup_conda_env.sh

Quickstart: train an agent in 2 minutes

export OPENAI_API_KEY=sk-...

# Up to three improvement iterations, LLM-as-judge
praisonai-train agents --input "Explain quantum entanglement to a 10-year-old" --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?"

Note: --iterations sets the maximum number of training loops. In LLM-as-judge mode, training stops early when any iteration scores ≥ 9.5 (excellent), so easy prompts may finish in a single iteration. Pass --no-early-stop to force all iterations, or --verbose to see when it stops.

Python API:

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

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

# 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.

Dataset tooling (generate + validate)

Build and quality-check instruction datasets — protocol-driven and YAML-configurable.

# Synthesize from a teacher LLM (recipe + diversity axes, JSON mode, dedup, resumable offsets)
praisonai-train generate --config generate.yaml
praisonai-train generate -r tamil -d gpt-4o -n 1000 -o data/tamil.jsonl

# Quality-check / filter (dedup, boilerplate & refusal, script purity, diversity metrics)
praisonai-train validate data/tamil.jsonl --out data/clean.jsonl

Add a language/domain by registering a Recipe, or a new QC rule by registering a RowCheck (see praisonai_train/data/), and they show up automatically.

Verify → export → train → re-verify (learn from real agent runs)

Beyond synthetic generation, you can fine-tune on the agent's own verified behaviour. Given a trials report (K scored attempts per case with captured trajectories), from-trials keeps the passing attempts and writes a trainer-ready dataset — closing the loop from verification to data generation.

# 1) export the passing attempts as a ShareGPT dataset (+ provenance sidecar)
praisonai-train data from-trials trials.json -o data/train.jsonl

# 2) fine-tune with the existing trainer, unchanged
praisonai-train llm --dataset data/train.jsonl

# 3) re-run the trials on the fine-tuned model and compare pass-rate to measure gain

from praisonai_train.data import export_trials

summary = export_trials(
    report, "data/train.jsonl",
    only_passed=True,     # rejection sampling on the verifier (default)
    frontier_only=True,   # skip saturated / zero-pass cases (default)
)
print(summary)  # written / skipped_failed / skipped_tool_runs / skipped_no_text / ...

Selection defaults: unscored attempts are never candidates; only_passed keeps verifier-passed attempts; frontier_only restricts to cases with 0 < pass_rate < 1 (saturated cases add near-duplicates of mastered behaviour, zero-pass cases have nothing to export); tool-using runs are excluded by default (--all, --include-saturated relax these). A {out}.jsonl.meta.json sidecar maps every line to its case id, attempt index and score, and emission order is deterministic → reproducible files. Add --qc to run rows through the same QC filter as validate, or --format alpaca for instruction/input/output.

Honest selection: only_passed is rejection sampling — it amplifies behaviour the agent already produces and inherits any bias in the scorer/judge that decided a run "passed". It cannot teach behaviour the agent never exhibited; treat it as reinforcing verified wins, not as an oracle.