kimeiga's picture
Initial upload
f710cff verified
|
Raw History Blame Contribute Delete
4.85 kB
---
pretty_name: base-board-eval threads
license: cc-by-4.0
language:
- en
size_categories:
- n<1K
tags:
- base-models
- llm-evaluation
- synthetic
- imageboard
- sampling
viewer: false
extra_gated_heading: "This dataset contains unfiltered model output"
extra_gated_prompt: >-
These threads were generated by base language models with no system prompt, no safety
tuning in play, and no human curation. Some posts contain slurs, sexual violence and
other disturbing material. Threads caught by an automatic filter have their posts
withheld, but the filter is a keyword match and missed a lot. Only request access if you
are studying model behavior and can handle this material.
extra_gated_button_content: "I understand, request access"
extra_gated_fields:
I understand this data contains offensive and disturbing text: checkbox
I will not republish this data without filtering it: checkbox
---
# base-board-eval threads
The raw generated threads behind
[studiokimeiga/base-board-eval](https://github.com/studiokimeiga/base-board-eval), an
evaluation of whether small **base** language models can write imageboard-style threads that
read like real people talking.
This is about 690 threads, plus 72 ablation threads, from 4–8B base checkpoints (gemma-4-E4B, Qwen3-4B-Base,
Olmo-3-1025-7B, Mistral-7B-v0.3, plus one instruct model as a comparison arm), run locally on
llama.cpp. The statistics in the repo's `FINDINGS.md` were computed from these files.
## Content warning
This is **unfiltered base-model output**. Base models given a mundane opener sometimes
produce slurs, sexual violence and self-harm content without being prompted to. The
`QUARANTINE.md` file in the code repo puts the rate of self-harm content at about 1 thread
in 36 in one run.
- Threads that matched the automatic filter have `"quarantined": true`. Everything after the
opener is replaced with a notice like `[31 posts withheld - auto-quarantined]`. The
`matched` field lists the terms that triggered it, so it contains slurs itself.
- One self-harm thread in `threads.json` (Mistral-7B, opener 4) was withheld by hand.
- Posts that did **not** match the filter are untouched. Several of them describe rape or
sexual violence. Treat every file as unreviewed.
## Files
Each file is the output of one script in the code repo. Run scripts are in `experiments/`,
and run 14 is `generate.py` at the top level.
| file | script | what it tested |
|---|---|---|
| `threads.v1.json`, `threads.v2.json`, `threads.v3names.json` | early `generate1`–`3` | First blind base-model comparisons (snapshots kept before `threads.json` was overwritten) |
| `threads.json` | `generate3.py` (last writer) | 6 openers × 6 threads per arm, across five Q4_K_M arms |
| `threads.warmup.json` | `generate4.py` | gemma-4-E4B base at Q8_0 alone, 16-post threads |
| `threads.run5.json` | `generate5.py` | Volume run, 142 threads: how often one base model produces a keeper |
| `threads.long.json` | `generate6.py` | Mixed three-model cast, 50-post threads |
| `threads.upd.json` | `generate7.py` | OP forced to post at fixed points with a short seed |
| `threads.iso.json` | `generate8.py` | Run 6 with only DRY changed |
| `threads.dry3.json` | `generate9.py` | DRY `allowed_length` 3 |
| `threads.xtc.json` | `generate10.py` | XTC on vs. off |
| `threads.fmt.json` | `generate11.py` | Nothing injected into any post |
| `threads.xtc2.json` | `generate12.py` | XTC on vs. off, repeated |
| `threads.sweep.json` | `generate13.py` | XTC strength sweep |
| `threads.final.json`, `threads.final-partial-t05.json` | `generate.py` (run 14) | Final config: paired XTC strength and DRY on/off |
| `threads.blind.json` | `generate15.py` | Blind three-arm coherence test |
| `threads.optag.json` | `generate16.py` | Marking the thread starter's posts with (OP) |
| `threads.memory.json` | `generate17.py` | Board memory: does a coinage carry over to a new thread? |
| `ablation.json`, `ablation2.json`, `ablation3.json` | `ablate1`–`3.py` | Sampler ablations for distribution drift |
| `queue.json`, `queue.key.json` | `make_queue.py` | Blind rating queue for run 5, and its answer key |
| `shortlist.json` | — | Shortlist of run-5 threads chosen before rating |
The schemas differ between runs because the generators changed between runs. Most `threads.*`
files have a top-level `threads` list. Each thread has its participants, the model behind each
voice, and `posts` with `who`, `ts`, `text` and, in later runs, `voice` and `model`.
To browse a file, put it next to the code repo's HTML viewers (`review.html`, `blind.html`,
`xtc.html`, ...) and serve the folder with `python -m http.server`.
## License
The threads are released under CC BY 4.0. They were generated by the models listed above.
Check each model's license for any terms on its outputs.