---
dataset_info:
  features:
  - name: model
    dtype: string
  - name: api_base
    dtype: string
  - name: max_links
    dtype: int32
  - name: max_tries
    dtype: int32
  - name: start_article
    dtype: string
  - name: destination_article
    dtype: string
  - name: seed
    dtype: int32
  - name: result
    dtype: string
  - name: steps
    sequence:
    - name: step
      dtype: int32
    - name: type
      dtype: string
    - name: current_article
      dtype: string
    - name: action
      dtype: string
    - name: observation
      dtype: string
    - name: prompt
      dtype: string
    - name: response
      dtype: string
  splits:
  - name: qwen3_14b
    num_bytes: 230982
    num_examples: 420
  download_size: 12637
  dataset_size: 230982
configs:
- config_name: default
  data_files:
  - split: qwen3_14b
    path: data/qwen3_14b-*
---
# Wikihop Experiment Results

Results generated from: `qwen3-final-results.json`

## Experiment Parameters
- **Number of Trials per Pair:** 1
- **Max Steps per Game:** 20
- **Agent Settings:**
  - Model: `hosted_vllm/Qwen/Qwen3-14B`
  - API Base: `http://ip-26-0-163-127:8000/v1/`
  - Max Links: 500
  - Max Tries: 3

## Summary Statistics
- **Total Runs:** 420
- **Win Rate:** 55.24%
- **Average Hops (on Wins):** 6.64