harshith0214/psse-agent-gpt-oss-20b-lora

PSSE Agent — GPT-OSS 20B LoRA (v2)

A LoRA adapter fine-tuned on GPT-OSS 20B for automated Power System State Estimation (PSSE) fault diagnosis.

Given a power system snapshot and WLS state estimation outputs (residuals r, normalized Lagrange multipliers λN), the model classifies the fault into one of four categories with high confidence.

Model Details

Property	Value
Base model	unsloth/gpt-oss-20b-unsloth-bnb-4bit
Architecture	GPT-OSS 20B (32-expert MoE)
Adapter type	LoRA (PEFT)
LoRA rank / alpha	64 / 64
Target modules	q/k/v/o/gate/up/down projections + MoE experts
Trainable parameters	31.85M / 20.9B (0.15%)
Training epochs	3
Final train loss	0.005
Final eval loss	0.005
Peak GPU memory	56 GB (H100 80GB)
Quantization	4-bit (QLoRA)

Task

The model acts as a power-system diagnostic agent using an agentic tool-call loop:

1. Calls wls_from_path on the provided snapshot
2. Inspects residuals r and Lagrange multipliers λN
3. Optionally calls correction tools (correct_measurements_from_path, correct_parameters_from_path, correct_topology_from_path)
4. Outputs a structured JSON verdict

Output Format

{
  "has_error": true,
  "error_family": "parameter_error",
  "suspect_location": {"line_row": 5, "from_bus": 3, "to_bus": 4},
  "recommended_tool": "correct_parameters_from_path",
  "confidence": 0.99
}

Error Classes

Class	Description
measurement_error	Bad sensor reading — concentrated large residuals at specific channels
parameter_error	Incorrect line parameters — large Lagrange multipliers on a specific branch
topology_error	Incorrect network topology — widespread residual pattern
no_error	System healthy — residuals within normal bounds

Evaluation Results (held-out test set, IEEE 14-bus)

Class	Precision	Recall	F1	Support
measurement_error	100.0%	100.0%	100.0%	75
no_error	100.0%	100.0%	100.0%	75
parameter_error	100.0%	100.0%	100.0%	75
topology_error	100.0%	100.0%	100.0%	71
macro avg	100.0%	100.0%	100.0%	296

• Overall accuracy: 296/296 (100%)
• Mean confidence: 0.9875
• Null predictions (parse failures): 0/296

Note: Evaluated on held-out samples from the IEEE 14-bus (case14) synthetic dataset. Out-of-distribution generalization to other grid topologies has not been evaluated.

Training Data

• Source: Synthetically generated power system diagnostic traces (case14, IEEE 14-bus system)
• Total samples: 1,978 (deduplicated on user snapshot)
• Split: 70% train (1,400) / 15% val (297) / 15% test (296)
• Split strategy: Stratified by error_family × tool_sequence, seed=42
• Format: GPT-OSS channel format (commentary{}/final{})

Usage

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="harshith0214/psse-agent-gpt-oss-20b-lora",
    max_seq_length=16384,
    dtype=None,
    load_in_4bit=True,
)
FastLanguageModel.for_inference(model)

Training Infrastructure

• Hardware: NVIDIA H100 80GB HBM3
• Framework: Unsloth + HuggingFace TRL (SFTTrainer)
• Optimizer: AdamW 8-bit, cosine LR schedule
• Learning rate: 1e-4, warmup 50 steps
• Batch size: 4 × 4 gradient accumulation = 16 effective

Limitations

• Trained and evaluated exclusively on the IEEE 14-bus (case14) synthetic test system
• Performance on larger or real-world grid topologies is untested
• Requires the GPT-OSS base model and Unsloth for inference