qwen35-hua-sdf-lora

Qwen3.5-4B Hua SDF LoRA

This is a PEFT LoRA adapter trained with Tinker on the Hua et al. synthetic document finetuning (SDF) corpus, rewritten only for Qwen-compatible assistant identity phrasing.

Provenance

• Base model: Qwen/Qwen3.5-4B
• Training system: Tinker
• Training run ID: qwen35_hua_sdf_full_v1
• Tinker training checkpoint: tinker://c98f23c8-8c09-5e84-8bd2-3c039422024d:train:0/weights/qwen35_hua_sdf_full_v1
• Tinker sampler checkpoint: tinker://c6f382d1-90d7-58ad-9d3c-484dd680b01c:train:0/sampler_weights/qwen35_hua_sdf_full_v1_sampler
• Dataset ID: ds_hua_sdf_qwen35_identity_v1
• Split ID: split_hua_sdf_qwen35_identity_v1_train_all
• Examples trained: 311083
• Rendered train tokens: 175717024
• Loss tokens: 174783775
• Truncated examples: 0
• LoRA rank: 64

Intended Use

This checkpoint is a research artifact for studying synthetic-document finetuning and eval-awareness-style behavior transfer. It is the SDF-stage organism checkpoint, not the final EI-replay checkpoint.

Loading

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_model = "Qwen/Qwen3.5-4B"
adapter = "Luxel/qwen35-hua-sdf-lora"

tokenizer = AutoTokenizer.from_pretrained(base_model)
model = AutoModelForCausalLM.from_pretrained(base_model, device_map="auto")
model = PeftModel.from_pretrained(model, adapter)

Limitations

This model was trained to reproduce a specific SDF stage for controlled research. Do not treat it as a generally improved Qwen checkpoint.