PlanGuard is trained to critique, repair, reject, or approve structured onchain action plans before wallet execution. This seed adapter is a proof-of-work artifact for the public build process, not a production wallet-safety model.

Base Model

mlx-community/Qwen2.5-0.5B-Instruct-bf16

Training Data

CodePit/OnchainPlanBench-Seed

Intended Behavior

PlanGuard should learn to:

choose valid tools from a manifest,
avoid wrong-chain and wrong-token plans,
reject unsupported tool requests,
repair unsafe draft plans,
add wallet-confirmation gates,
minimize wallet context sent to inference,
recommend private RPC or private inference when appropriate,
reject laundering, sanctions-evasion, stolen-funds, or compliance-evasion requests,
output strict JSON plans that downstream validators can check.

Important Limitations

This is a seed preview adapter.
It is not a production safety model.
It does not authorize transactions.
It does not provide financial, legal, or compliance advice.
A model is not considered officially improved until CodePit's verifier scores it on held-out benchmark tasks.

Local Use

Install MLX-LM and load this adapter with the matching base model:

bash

pip install mlx-lm
mlx_lm.generate \
  --model mlx-community/Qwen2.5-0.5B-Instruct-bf16 \
  --adapter-path . \
  --prompt '<your PlanGuard prompt>'

Evaluation

See:

TRAINING_REPORT.md for the public summary of what was achieved and learned.
baseline-comparison.json for the base-vs-adapter validation metrics.
base-generations.jsonl and adapter-generations.jsonl for raw generation outputs used by the lightweight evaluator.
eval-summary.json for the MLX masked test loss/perplexity.

CodePit/PlanGuard-0.2-Seed-LoRA

CodePit PlanGuard 0.2 Seed LoRA