A small language model fine-tuned specifically for photography education — covering composition, exposure, lighting, color theory, post-processing, and camera equipment.

Final Training Results

Metric	Value
Best eval loss	2.114
Best eval token accuracy	53.6%
Final train loss	1.94
Training epochs	~2 of 3 (stopped early — eval loss plateaued)
Total training steps	~2,200

Model Details

Base Model: Qwen/Qwen3-4B — Apache 2.0
Fine-tuning Method: QLoRA (4-bit NF4 quantization + LoRA r=64, all-linear)
Training Data: 15,907 photography Q&A pairs from Photo Stack Exchange + curated synthetic examples
Trainable Parameters: 132.1M / 2,337.9M (5.65%)
Hardware: NVIDIA L4 (24GB VRAM)

What It Knows

Topic	Coverage
Exposure Triangle	Aperture, shutter speed, ISO, metering, histograms
Composition	Rule of thirds, leading lines, framing, negative space, golden ratio
Lighting	Natural light, studio lighting, flash, modifiers, portrait patterns
Color	White balance, color temperature, HSL, color grading, split toning
Post-Processing	RAW development, Lightroom/Photoshop workflows, HDR, focus stacking
Equipment	Lenses, camera bodies, filters, tripods, lighting gear
Genres	Portrait, landscape, street, wildlife, macro, astrophotography

Usage (LoRA Adapter)

This repo contains the LoRA adapter. Load it on top of the base model:

python

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

# Load base model
base_model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen3-4B",
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("Builder-Neekhil/photography-fundamentals-qwen3-4b")

# Load LoRA adapter
model = PeftModel.from_pretrained(base_model, "Builder-Neekhil/photography-fundamentals-qwen3-4b")

# Generate
messages = [
    {"role": "system", "content": "You are an expert photography instructor."},
    {"role": "user", "content": "How do I use the exposure triangle for sunset photography?"}
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512, temperature=0.7, top_p=0.8)
print(tokenizer.decode(outputs[0][inputs.input_ids.shape[-1]:], skip_special_tokens=True))

Merge adapter for standalone deployment

python

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-4B", torch_dtype="auto")
model = PeftModel.from_pretrained(base_model, "Builder-Neekhil/photography-fundamentals-qwen3-4b")
merged = model.merge_and_unload()
merged.save_pretrained("photography-slm-merged")
AutoTokenizer.from_pretrained("Builder-Neekhil/photography-fundamentals-qwen3-4b").save_pretrained("photography-slm-merged")

Training Config

Parameter	Value
Learning rate	2e-4 (cosine decay)
Warmup steps	150
Effective batch size	16 (2 × 8 grad accum)
Max sequence length	2048
Weight decay	0.01
LoRA rank (r)	64
LoRA alpha	32
LoRA dropout	0.05
Target modules	all-linear (q,k,v,o,gate,up,down)
Quantization	4-bit NF4, double quant
Early stopping	Manual at epoch ~2 (eval loss plateau)

Dataset

Built from Photo Stack Exchange via HuggingFaceH4/stack-exchange-preferences (CC-BY-SA 4.0), filtered for photography topics with pm_score ≥ 1. Supplemented with curated synthetic examples.

Dataset: Builder-Neekhil/photography-fundamentals-sft (15,907 train / 838 test)

Limitations

Text-only — cannot analyze actual photographs
Knowledge reflects Stack Exchange community consensus
May occasionally generate HTML artifacts from source data
Best suited as a photography tutor/reference, not for creative writing

License

Apache 2.0 (inherited from Qwen3-4B). Training data under CC-BY-SA 4.0.