Balrog
Harbor: pendingTracked by Epoch AI's Benchmarking Hub (results collected from external sources).
Incomplete29%Knowledge & Reasoning
- Samples
- N/A
- Released
- 2024
- Scores
- 39
PFaithBench
Pairwise Faithfulness Benchmark
Evaluates LLM faithfulness by testing if models switch between answering and abstaining for the same question under sufficient vs. insufficient contexts.
How to run
ltb · pfaithbench
$ npm i -g lettokenburn-cli$ ltb bench info pfaithbench
External harness. No Harbor adapter has been generated for this benchmark yet. `ltb bench info` prints the known links.
Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs
Scores
11 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| Llama-3.2 1B | 48.8 | N/A | Independent | medium | Oct 6, 2026 | arxiv.org |
| Llama-3.1-8B-Instruct | 48.7 | N/A | Independent | medium | Oct 6, 2026 | arxiv.org |
| Qwen-3-0.6B | 48.3 | N/A | Independent | medium | Oct 6, 2026 | arxiv.org |
| Qwen3 1.7B | 48.2 | N/A | Independent | medium | Oct 6, 2026 | arxiv.org |
| Qwen3-32B | 33.8 | N/A | Independent | medium | Oct 6, 2026 | arxiv.org |
| Llama-3.3-70B-Instruct | 32.6 | N/A | Independent | medium | Oct 6, 2026 | arxiv.org |
| Llama-3.1-70B-Instruct | 29.2 | N/A | Independent | medium | Oct 6, 2026 | arxiv.org |
| Qwen3 4B | 27.3 | N/A | Independent | medium | Oct 6, 2026 | arxiv.org |
| Qwen3 14B | 26.2 | N/A | Independent | medium | Oct 6, 2026 | arxiv.org |
| Llama-3.2 3B | 22.7 | N/A | Independent | medium | Oct 6, 2026 | arxiv.org |
| Qwen3 8B | 18.4 | N/A | Independent | medium | Oct 6, 2026 | arxiv.org |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
Reproduced by the community
All runsLoading community runs…
About
PFaithBench constructs pairs of sufficient and insufficient contexts for the same question from multi-hop, single-hop, and counterfactual QA datasets. It measures context-sensitive behavior rather than isolated instance accuracy.
Scoring focuses on identifying over-answering (fabricating responses with insufficient context) and over-abstention (refusing to answer with sufficient context). Results show most models bias toward over-answering.