OneReason-0.8B-lora-expA

OneReason-0.8B-LoRA-ExpA

LoRA adapter fine-tuned from OpenOneRec/OneReason-0.8B-pretrain-competition on the Kuaishou Explorer LLM-Rec Challenge dataset.

Model Details

• Base Model: OpenOneRec/OneReason-0.8B-pretrain-competition (Qwen3-0.8B)
• Fine-tuning Method: LoRA (rank=16, alpha=32)
• Target Modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• Training Data: Kuaishou LLM-Rec Challenge competition dataset (32K samples)
• Training Steps: 4,660 / 11,571 (~40% of 3 epochs)
• Hardware: NVIDIA RTX 4090 (47GB VRAM)
• Training Time: ~2h 41min

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model = "OpenOneRec/OneReason-0.8B-pretrain-competition"
model = AutoModelForCausalLM.from_pretrained(
    base_model,
    trust_remote_code=True,
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained(base_model, trust_remote_code=True)

# Load LoRA adapter
model = PeftModel.from_pretrained(model, "dfdu233/OneReason-0.8B-lora-expA")

# Example inference
prompt = "<|prod_begin|><s_a_1183><s_b_746><s_c_5290>，这个商品卖的是什么？ /no_think"
messages = [{"role": "user", "content": prompt}]
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, do_sample=True, top_p=0.95, temperature=0.7)
print(tokenizer.decode(outputs[0][len(inputs.input_ids[0]):], skip_special_tokens=True))

Training Details

• Loss: 3.023 → 1.337 (56% reduction over 4660 steps)
• Validation Loss: 1.373 (decreasing, no overfitting)
• Cutoff Length: 8192
• Batch Size: 1 (gradient accumulation 8 = effective batch 8)
• Learning Rate: 3e-5 (cosine schedule)
• Optimizer: paged_adamw_8bit
• Context Length: 8192 tokens
• Flash Attention: fa2

Training Data Distribution

Framework versions

• PEFT 0.18.1
• PyTorch 2.4.0
• Transformers 5.6.0
• Flash-Attn 2.6.3