metadata

license: apache-2.0
base_model: thomasjhuang/qwen2-sft-warmup
tags:
  - reinforcement-learning
  - rloo
  - countdown-math
  - qwen2
language:
  - en
pipeline_tag: text-generation

Qwen2 RLOO Countdown (Step 250)

This model is a Qwen2-based language model fine-tuned using RLOO (Rank-order Learning with Localized Objectives) on countdown math problems.

Training Details

Base Model: thomasjhuang/qwen2-sft-warmup
Method: RLOO (Reinforcement Learning from Human Feedback)
Dataset: Jiayi-Pan/Countdown-Tasks-3to4
Training Steps: 250 optimizer steps
Learning Rate: 3e-6
Temperature: 0.1
Batch Size: 2
K Samples: 8

Key Fixes Applied

Prompt Format: Updated to match SFT evaluation format with detailed instructions
Token Length: Increased to 250 tokens for complete reasoning
Temperature: Reduced to 0.1 for more deterministic generation
Extraction: Fixed to work with vLLM outputs

Performance

During training at step 250, the model achieved:

Average rewards ranging from 0.05 to 0.50 across batches
Successful generation of proper <think> and <answer> tags
Correct solutions to various countdown math problems

Usage

from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("thomasjhuang/qwen2-rloo-countdown-step250")
model = AutoModelForCausalLM.from_pretrained("thomasjhuang/qwen2-rloo-countdown-step250")

prompt = '''Using the numbers [8, 16, 80], create an equation that equals 72. You can use basic arithmetic operations (+, -, *, /) and each number can only be used once. Show your work in <think> </think> tags. And return the final answer in <answer> </answer> tags, for example <answer> (1 + 2) / 3 </answer>.'''

inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_length=300, temperature=0.1)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)

Training Progress

This checkpoint represents an intermediate state in RLOO training where:

The model learned to follow the correct prompt format
Success rates improved from 0% to 10-50% on various problems
The model generates structured reasoning in <think> tags
Solutions are properly formatted in <answer> tags

For the latest checkpoint, see: thomasjhuang/qwen2-rloo-countdown-final