qwen3-1.8b-semantic-ids/README.md

language, tags, license, base_model, datasets

language	tags	license	base_model	datasets
en	semantic-ids recommendation generative-retrieval qwen3 fine-tuned	apache-2.0	Qwen/Qwen3-1.7B	amazon-pet-supplies

Qwen3-1.8B Fine-tuned for Semantic ID Recommendation

Overview

Qwen3-1.8B fine-tuned for generative product recommendation via hierarchical semantic identifiers. The model generates 4-level Semantic IDs (<|sid_start|><|A#|><|B#|><|C#|><|D#|><|sid_end|>) given product descriptions, purchase histories, or co-purchase contexts.

This is the smaller model in a controlled comparison experiment (1.8B vs 8B) conducted under identical training conditions.

Training

Stage 1: Vocabulary Expansion

• Added 1,027 special tokens (3 structural + 4×256 codebook tokens)
• Trained only embedding matrices (0.3% of parameters)
• 2,000 steps, LR 1×10⁻³, batch 64

Stage 2: Full Fine-tuning

• Dataset: 4,719,994 instruction-formatted conversations (Amazon Pet Supplies)
• Task types: text→SID, sequential recommendation, co-purchase prediction
• Optimizer: AdamW 8-bit, LR 2×10⁻⁵, cosine with min LR (0.2×peak)
• Warmup: 3%, weight decay 0.01
• Batch: 64 × 2 = 128 effective, 3 epochs
• Techniques: Custom instruction masking, greedy sequence packing (~3× throughput)
• Hardware: NVIDIA H100 80GB (vast.ai)

Results

Hierarchical SID prediction accuracy (A-level match, greedy decoding):

Task	Accuracy
Text → SID	59.9%
Sequential recommendation	7.0%
Co-purchase prediction	5.5%

Evaluation: 3,000 samples per task, 11 task types.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("kalistratov/qwen3-1.8b-semantic-ids")
tokenizer = AutoTokenizer.from_pretrained("kalistratov/qwen3-1.8b-semantic-ids")

Citation

Master's thesis, Moscow Institute of Physics and Technology (MIPT), 2026.

References

1. Y. Sun et al. "OpenOneRec," arXiv:2502.18851, 2025.
2. J. Liu et al. "PLUM," arXiv:2406.12346, 2024.
3. E. Yan. "semantic-ids-llm," GitHub, 2024.