Downloads · 30 days
11
7% of all-time downloads
beachcities/qwen3-4b-sft-v5h-hybrid-merged
qwen3-4b-sft-v5h-hybrid-merged is a text generation model from beachcities. Use it when you need the model to write or continue text. The card lists the license as apache-2.0.
- Model Name: beachcities/qwen3-4b-sft-v5h-hybrid-merged - Base Model: Qwen/Qwen2.5-3B-Instruct (approx. 4B parameters) - Model Type: Causal Language Model (Fine-tuned via SFT & DPO/Hybrid optimization) - Language(s):…
Downloads · 30 days
11
7% of all-time downloads
All-time downloads
159
Public
Parameters
4B
8.1 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors8 GB · 100%
From the Hugging Face model README
The dataset was carefully engineered to balance "Offense" (deep structural learning) and "Defense" (preventing format hallucination), building upon a baseline of approximately 12,686 samples.
The model was trained in two distinct stages: a robust SFT phase followed by a DPO alignment phase. (Note: Rank-Stabilized LoRA was not utilized in this specific run.)
Stage 1: Supervised Fine-Tuning (SFT)
r=64, lora_alpha=64, use_rslora=False, target_modules=["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"]learning_rate: 5e-5lr_scheduler_type: cosinenum_train_epochs: 2per_device_train_batch_size: 2gradient_accumulation_steps: 8 (Effective Batch Size = 16)warmup_ratio: 0.1Stage 2: Direct Preference Optimization (DPO)
r=8, lora_alpha=16, use_rslora=Falselearning_rate: 1e-7beta: 0.1num_train_epochs: 1per_device_train_batch_size: 2gradient_accumulation_steps: 4 (Effective Batch Size = 8)optimizer: adamw_8bitThis model provides a critical baseline for exploring the internal mechanics of formatting vs. reasoning in LLMs. Future research based on this artifact will focus on the following phenomena:
= { }), or wraps complex XML structures inside Markdown <code> blocks. This highlights the inherent limitation of SFT in correcting structural "laziness" embedded during pre-training.<think>...</think>) frequently interferes with the structural integrity of subsequent strict formatting (especially in XML and TOML).<think></think> block to bypass reasoning overhead while preserving the structural precision of the output.