Qwen3B-s1k-ssc-tiny-sft-exp0

Full fine-tuned Qwen/Qwen2.5-3B-Instruct on the s1k-compressed dataset using ssc_reasoning_trace reasoning traces.

Model Description

This model is a fully fine-tuned version of Qwen2.5-0.5B-Instruct (all 494M parameters trained, not LoRA) for mathematical reasoning tasks.

Training Details

• Base Model: Qwen/Qwen2.5-3B-Instruct
• Training Type: Full SFT (all parameters updated)
• Dataset: robin-linzmayer/s1k-compressed
• Reasoning Trace: ssc_reasoning_trace
• Training Examples: 99
• Validation Examples: 6
• Max Sequence Length: 4096
• Final Training Loss: 0.4768
• Final Eval Loss: 0.49350830912590027

Training Configuration

learning_rate: 1e-05
num_train_epochs: 3
per_device_train_batch_size: 2
gradient_accumulation_steps: 4

Usage

from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("robin-linzmayer/Qwen3B-s1k-ssc-tiny-sft-exp0")
model = AutoModelForCausalLM.from_pretrained("robin-linzmayer/Qwen3B-s1k-ssc-tiny-sft-exp0")

# Example usage for reasoning tasks
prompt = "Solve this math problem: What is 2 + 2?"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=512)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)

Training Data

This model was trained on the s1k-compressed dataset, which contains compressed reasoning traces for mathematical problem-solving.