You can deploy this model and start earning money today!

siddadatrao/
fragrance-cpt-qwen3-32b-v2

Name: fragrance-cpt-qwen3-32b-v2
Availability: InStock

continued-pretraining

Views

No views yet

Fragrance CPT Qwen3 32B V2

Fragrance V2 is a continued-pretraining PEFT adapter for unsloth/qwen3-32b-bnb-4bit. It continues the V1 adapter siddadatrao/fragrance-cpt-qwen3-32b for an additional 500 optimizer steps on a packed corpus of fragrance discussions.

This model is intended for experimentation with fragrance-domain language and consumer discourse. It is not a calibrated survey simulator: probabilities or repeated samples from the model must not be interpreted as population survey percentages without separate distribution-alignment training and held-out validation.

Training data

The source corpus contains 5,926,634 fragrance-discussion rows. These were packed into 170,242 sequences with a target length of 2,048 tokens. The packing manifest estimates 325,320,274 source tokens in total.

The corpus is drawn from online fragrance discussion and is expected to overrepresent enthusiasts and active online participants. It should not be treated as a representative sample of fragrance buyers.

V2 continuation configuration

Starting adapter: siddadatrao/fragrance-cpt-qwen3-32b
Base model: unsloth/qwen3-32b-bnb-4bit
Objective: causal next-token prediction
Additional optimizer steps: 500
Maximum sequence length: 2,048
Per-device batch size: 2
Gradient accumulation: 8
Effective batch size: 16 sequences
Learning rate: 1e-5
Embedding learning rate: 1e-6
Warmup ratio: 0.03
Validation ratio: 0.0004 (69 packed sequences)
Seed: 4407
LoRA rank: 64
LoRA alpha: 32
Rank-stabilized LoRA: enabled
LoRA dropout: 0
Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj, embed_tokens, and lm_head

Usage

python

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

adapter_id = "siddadatrao/fragrance-cpt-qwen3-32b-v2"
base_id = "unsloth/qwen3-32b-bnb-4bit"

tokenizer = AutoTokenizer.from_pretrained(adapter_id)
base = AutoModelForCausalLM.from_pretrained(
    base_id,
    device_map="auto",
    dtype=torch.bfloat16,
)
model = PeftModel.from_pretrained(base, adapter_id)
model.eval()

The adapter is approximately 5 GB because the checkpoint includes trained embedding weights in addition to the rank-64 LoRA tensors.

Limitations

The model reflects patterns in its online training corpus, not a statistically representative population.
It was not trained against survey response distributions.
First-token option probabilities are not automatically calibrated to human response rates.
Domain adaptation can change general-language behavior; evaluate against both in-domain and general-text controls before deployment.
Generated claims about products, safety, health, or consumer statistics require independent verification.