Polyvalent TEP Report-Writer Adapter (Step 447)

PEFT LoRA adapter for industrial root-cause analysis (RCA) and diagnostic report generation on the Tennessee Eastman Process (TEP) benchmark.

Model Description

  • Adapter Name: internvl3_2b_pristine_rca
  • Base Model: OpenGVLab/InternVL3-2B (pinned at revision 899155015275a9b7338c7f4677e19c784e0e5a21)
  • Parameters: 18,464,768 trainable LoRA parameters
  • Checkpoint Save: Final step 447 (epoch 3.0)

Intended Use

  • Generates structured engineering root-cause analysis reports from tabular process telemetry, fault symptom windows, and control loop disturbances.
  • Answers plant operator follow-up queries regarding valve saturation, pressure spikes, reactor temperature drift, and shutdown risk.

Training Details

  • Architecture: LoRA rank $r=16$, $\alpha=32$, dropout $0.05$ applied to q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj.
  • Training Data: 2,376 training examples covering all 22 TEP fault classes distilled from qwen2.5:32b.
  • Optimization: Cosine learning rate schedule ($2\times 10^{-4}$), warmup ratio 0.03, bf16 mixed precision, effective batch size 16.

Evaluation and Validation

Evaluated across in-distribution full-coverage fault test suites:

  • Grounding Rate: 100% sensor variable verification against simulator truth.
  • Valid Syntax: 100% compliant structured report JSON/Markdown format.
  • Notice: Scores measured on the v0 distillation set reflect in-distribution performance across all 22 known fault dynamics.

How to Load

import torch
from transformers import AutoModel, AutoTokenizer
from peft import PeftModel

BASE_MODEL = "OpenGVLab/InternVL3-2B"
BASE_REVISION = "899155015275a9b7338c7f4677e19c784e0e5a21"

tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL, revision=BASE_REVISION, trust_remote_code=True)
base_model = AutoModel.from_pretrained(
    BASE_MODEL,
    revision=BASE_REVISION,
    torch_dtype=torch.bfloat16,
    trust_remote_code=True
).eval()

model = PeftModel.from_pretrained(base_model, "Pushkarrrr15151/PolyVL-tep-report-lora").eval()

Licence

  • Model weights: Apache-2.0
  • Base model OpenGVLab/InternVL3-2B: Apache-2.0
  • Distillation teacher (qwen2.5:32b): Apache-2.0

Citation

@article{polyvalent2026,
  title={Polyvalent: Single-Base Multi-Specialist Architecture with Dynamic Adapter Residency},
  author={Polyvalent Research Team},
  year={2026}
}
Downloads last month
17
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for Pushkarrrr15151/PolyVL-tep-report-lora

Dataset used to train Pushkarrrr15151/PolyVL-tep-report-lora