Modelpublic

adarrshDev/ornith-1.5-9b-sft-r16

sourceHugging Facemitupdated 24d agoView on Hugging Face

0likes43downloads

Model Card

Ornith-1.5-9B — IEC 61131-3 Structured Text SFT (LoRA r=16)

A QLoRA fine-tuned LoRA adapter for `ornith-ai/Ornith-1.5-9B` (Qwen3.5 architecture) specialised in generating, refactoring, and reasoning about IEC 61131-3 Structured Text (ST) industrial control logic (OSCAT library standards, Siemens SCL dialects, and Plant-01 safety interlocks).

🔬 Model Details

Property	Value
Base Model	`ornith-ai/Ornith-1.5-9B`
Base Architecture	Qwen3.5 Dense (`qwen3_5`), Multimodal Reasoner
PEFT Method	QLoRA (4-bit NF4 quantized base + Float32/BFloat16 LoRA adapters)
LoRA Rank ($r$)	16
LoRA Alpha ($\alpha$)	32 ($\alpha/r = 2.0$)
LoRA Dropout	0.0
Target Modules	Attention & MLP projections (`q_proj`, `k_proj`, `v_proj`, `o_proj`, `gate_proj`, `up_proj`, `down_proj`)
Trainable Parameters	29,097,984 / 9,438,911,728 (0.31%)
Context Length	2,048 tokens (Base supports up to 262K)
Training Hardware	NVIDIA Tesla T4 (16 GB VRAM) on Google Cloud
Training Duration	446.1 seconds
Framework	Unsloth 2026.8.22 + TRL + HuggingFace PEFT
Base License	MIT

📊 Training Convergence & Loss Curve

Trained for 3 full epochs (9 optimization steps with effective batch size 8):

Step	Epoch	Training Loss	Learning Rate	Grad Norm
1/9	0.4	2.614	0.00e+00 (Warmup)	3.655
2/9	0.8	2.683	2.00e-04	4.080
3/9	1.0	2.691	1.92e-04	2.888
4/9	1.4	1.896	1.71e-04	1.636
5/9	1.8	1.646	1.38e-04	1.461
6/9	2.0	1.312	1.00e-04	1.519
7/9	2.4	1.228	6.17e-05	1.365
8/9	2.8	1.082	2.93e-05	1.558
9/9	3.0	1.105	7.61e-06	1.756

—Initial Loss: 2.614
—Final Converged Loss: 1.105
—Mean Loss across Run: 1.806

📦 Training Data Specification

—Domain: IEC 61131-3 Structured Text (ST) / Siemens SCL for PLC and DCS industrial automation.
—Corpus Structure: OSCAT basic automation library blocks, Plant-01 twin interlocks, and safety refusal prompts.
—Format: ChatML / OpenAI JSONL multi-turn format (messages schema).

💻 How to Load and Use with Transformers & PEFT

python

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

base_model_id = "ornith-ai/Ornith-1.5-9B"
adapter_id = "adarrshDev/ornith-1.5-9b-sft-r16"

# 1. Load Tokenizer & Base Model
tokenizer = AutoTokenizer.from_pretrained(adapter_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    torch_dtype=torch.float16,
    device_map="auto",
    trust_remote_code=True
)

# 2. Attach the Fine-Tuned Ornith LoRA Adapter
model = PeftModel.from_pretrained(model, adapter_id)
model.eval()

# 3. Generate Structured Text logic
messages = [
    {"role": "user", "content": "Write an IEC 61131-3 Structured Text FUNCTION_BLOCK for a debounced digital input with configurable ON/OFF delay timers."}
]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

with torch.no_grad():
    outputs = model.generate(**inputs, max_new_tokens=512, temperature=0.2)
print(tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))

⚖️ License & Governance

This adapter is distributed under the MIT License, consistent with the `ornith-ai/Ornith-1.5-9B` upstream base model.