stationhouse's picture Cerru02's picture
[experiment] [request for compute] Experiment: conspiracy versus ordinary falsehood training (#444)
dead516
|
Raw History Blame Contribute Delete
2.14 kB
# Completed experiments
Experiments that have been run, with their results. Each one starts as an
[experiment request](../experiment_requests/) and is moved here by a maintainer once the run
is done and the results are merged.
Experiments are grouped in sub-folders by the task they belong to. A request that was filed
under a task folder moves to the folder with the same name here:
| Task folder | What it covers |
|---|---|
| [`judge_reliability/`](./judge_reliability/) | Choosing the judge: judge baselines, Clef and decision-model runs, the validity set v1 and its labels, chat-judge comparisons (#50, #113, #122, #126, #203, #208); the validity-set diagnosis, label review, rubric experiments (six rules, endorsement rules) and the interactive judge explorer |
| [`judge_injection/`](./judge_injection/) | Prompt-injection fixtures and detector studies (PromptShield, Prompt Guard 2, Kev/Laya), in `experiments/` and `reports/` (#111, #173) |
| [`stage2_pilot/`](./stage2_pilot/) | Stage 2 evidence-ledger pilot: rendering hand-written evidence ledgers into questions |
| [`conspiracybench/`](./conspiracybench/) | Model-behaviour experiments on the benchmark itself: the evidence-focused system instruction (#441) |
| [`evals/`](./evals/) | Benchmark evaluation runs: the Phase 2 baseline suite (#428), Qwen3.5 4B with thinking off (#445), and the 4chan /x/ LoRA re-evaluation (#436) |
| [`hillclimb/`](./hillclimb/) | Fine-tuning transfer experiments: math sycophancy transfer to ConspiracyBench (#442); conspiracy versus ordinary falsehood training (#444) |
```
completed_experiments/<task>/<short-slug>/
EXPERIMENT.md the original request, unchanged
RESULTS.md what was run (models, versions, hardware, date), the numbers, and what they mean
... code, configs and the raw outputs or a link to them
```
The three folders above were moved here from `eval/judge_reliability/`, `eval/judge_injection/`
and `stage2_pilot/`, with their own layouts kept as they were. Recorded results (JSON, CSV,
HTML, logs and source archives) were not edited, so paths written inside them still use the
old locations.