README.md

---
license: cc-by-4.0
task_categories:
- text-generation
language:
- en
pretty_name: Community Alignment Modified Next-Human Followups
configs:
- config_name: default
  data_files:
  - split: source_examples
    path: data/source_examples-*
  - split: fewshot_demonstrations
    path: data/fewshot_demonstrations-*
  - split: validation
    path: data/validation-*
---

# Community Alignment Modified: next-human followups

This is a deterministic next-human-turn view of
[`facebook/community-alignment-dataset`](https://huggingface.co/datasets/facebook/community-alignment-dataset) at pinned
revision `97343c7f6399fcbea430ed0f37c1768281a78d56`. It contains 2,514 eligible
conversations from 90,256 source rows.

Five fixed rows are published only as `fewshot_demonstrations`. Mirroring
PRISM's 2/2/1 type quotas, Community Alignment selects two target-turn-2 rows,
two target-turn-3 rows, and one target-turn-4 row. Targets contain 30--200 words
and each prompt+target pair uses at most 256 MARIN tokens. Selection within each
stratum is hash-randomized with seed `42`; selected rows are removed
from the query pool rather than rerolled after inspection. The remaining
2,509 disjoint rows are published as `source_examples`.
Finetuning and validation prepend the five fixed `fewshot_demonstrations`
prompt+target demonstrations to each query using the same delimiter and task
instruction as lm-eval.

The published `validation` split performs that formatting once during dataset
construction. First, the shortest random-order prefix of complete source
examples reaching
650,000
source tokens is selected without counting demonstrations. It contains
650,269 source tokens.
Then the same five demonstrations are prepended to every selected query,
yielding 3,194,395
processed tokens after also counting one EOS boundary per document. In this
split, `prompt` stores the exact lm-eval context without the held-out target,
while `text` appends the target for full-document causal-LM validation loss.
The original query-only values remain in `source_examples` under the same
`conversation_id`. Runtime evaluation does not resample rows or reconstruct
demonstrations.

`source_examples` contains complete, source-derived next-human examples before
the five-shot prefix is added. Each `text` value contains the task instruction,
every available context turn used at the prediction boundary, and the genuine
human target; it is not a prompt-only or fragmented representation.

Each source conversation contributes at most one example. `target` is the
deepest available human follow-up that has at least one complete preceding
human/preferred-assistant exchange. Candidate assistant answers to the target
turn are excluded, preventing target-answer leakage.

## Next-human-turn representation

- `prompt` begins exactly with `Conversation:` and ends exactly with `Human: `.
- `target` is the source human follow-up, retained byte-for-byte.
- Earlier human prompts and preferred assistant responses are retained
  byte-for-byte and rendered with `Human:` / `Assistant:` labels.
- `text` is `Continue the conversation by writing exactly one next message from the human participant. Return only that message; do not include a speaker label, explanation, or additional turn.` followed by `prompt + target`; causal-LM
  loss can be applied to the entire field.
- `messages`, `training_messages`, preference keys, source identity, language,
  and annotator metadata are included for auditing.

No source utterance is paraphrased, trimmed, or normalized. The transformation
only chooses the prediction boundary and adds explicit speaker labels and an
open human cue for next-human-turn validation.

## PRISM-compatible filtering

- Only source rows whose `assigned_lang` is English and whose complete combined
  conversation is classified as English by pinned `langid==1.1.6` are published.
- Exact normalized model-facing conversations are unique. Grouping uses Unicode
  NFKC, case-folding, and collapsed whitespace. One source row is selected at
  random per duplicate group with seed `42`;
  group ordering uses an independent seeded hash so repeated source records do
  not receive extra sampling probability.
- `prompt` is at most 400 tokens.
- `target` is at most 256 tokens.
- Complete task-formatted `text` is at most 2048 tokens.
- Lengths use [`marin-community/isoflop-1ep20-d2304-L23-B128-dclm-default-4e21eb`](https://huggingface.co/marin-community/isoflop-1ep20-d2304-L23-B128-dclm-default-4e21eb)
  at `c932b4da5f56c721082e1a678ecfb8c0f4e9d4c7` with `add_special_tokens=False`, matching
  PRISM Trial 3's tokenizer and directly applicable per-row caps.
- No demographic, language, or conversation-type balancing is performed.

The upstream dataset is released under CC BY 4.0. Users should consult its
dataset card for collection details, intended use, and limitations.