README.md

---
language:
- en
license: unknown
size_categories:
- 1K<n<10K
configs:
- config_name: structured_WIP
  data_files:
  - split: train
    path: "tildes.parquet"
---

<div style="
  display: inline-block;
  width: fit-content;
  max-width: 100%;
  background: #0b0f14;
  border: 1px solid #223044;
  border-left: 5px solid #20c997;
  border-radius: 8px;
  padding: 14px 18px;
  margin-bottom: 20px;
  color: #e6edf3;
  font-family: ui-monospace, SFMono-Regular, Menlo, Monaco, Consolas, 'Liberation Mono', monospace;
  box-shadow: 0 8px 24px rgba(15, 23, 42, 0.18);
">
  <p style="
    margin: 0 0 6px 0;
    color: #8aa2bd;
    font-size: 12px;
    line-height: 1.4;
    text-transform: uppercase;
  ">codex dataset card</p>
  <p style="
    margin: 0;
    color: #e6edf3;
    font-size: 15px;
    font-weight: 650;
    line-height: 1.45;
  "><span style="color:#20c997;">&gt;</span> generated by <span style="color:#ffffff;">Codex</span><span style="color:#20c997;">_</span></p>
</div>

# Dataset Card for tildes-structured_WIP

## Dataset Summary

This is a partial, structurally validated supervised fine-tuning dataset derived
from the Tildes thread-only curation pipeline. It was generated from the first
3,884 checkpointed source topics of the `full_v1` run and contains 2,577 usable
training records after root curation, input-author subtree scrubbing, and
post-scrub trainability checks.

The intended training shape is:

- `input`: a source-only, user-like root prompt.
- `discussion.nodes`: a multi-author tree/forest of source discussion posts.
- `metadata`: provenance, curation, and audit fields not intended as model
  input.

There is no final-answer or `output` target. The intended SFT task is
`input -> discussion`.

This preflight slice is suitable for testing training templates, data loaders,
upload tooling, and downstream dataset consumers. It should not be treated as
the final `full_v1` dataset.

## Dataset Details

- **Language:** English
- **Source website:** https://tildes.net
- **Source crawl date:** 2026-05-03
- **Derived split:** `structured_preflight_3884`
- **Primary data file:** `tildes.parquet`
- **Companion source JSONL:** `tildes.jsonl`
- **Records:** 2,577
- **Discussion nodes:** 26,789
- **Blank discussion nodes:** 5
- **Rows with `output` field:** 0
- **Rows with `u0` in `discussion.nodes`:** 0

Root policy counts:

| Root policy | Records |
|---|---:|
| `topic_root` | 903 |
| `standalone_top_level` | 1,178 |
| `topic_context_top_level` | 496 |

Split status counts:

| Split status | Records |
|---|---:|
| `original` | 903 |
| `split` | 1,674 |

Text volume:

| Metric | Value |
|---|---:|
| Input characters | 2,478,840 |
| Input nonblank characters | 2,034,051 |
| Discussion node characters | 16,136,506 |
| Discussion node nonblank characters | 13,238,627 |
| Minimum discussion nodes per record | 1 |
| Maximum discussion nodes per record | 275 |

Parquet conversion:

| Metric | Value |
|---|---:|
| Parquet file size | 8.0 MB |
| Rows | 2,577 |
| Row groups | 1 |
| Max uncompressed row group size | 18.5 MB |
| Compression | ZSTD |

## Dataset Structure

Each row has the following top-level fields:

| Field | Type | Description |
|---|---|---|
| `input` | string | Source-only prompt intended to read like something a person might ask an assistant |
| `discussion.nodes` | list<object> | Source discussion tree/forest after input-author subtree scrubbing |
| `metadata` | object | Source, curation, filtering, and audit metadata |

Each discussion node has:

| Field | Type | Description |
|---|---|---|
| `id` | string | Per-example node ID, reindexed as `n0`, `n1`, etc. |
| `parent` | string or null | Parent node ID inside the same example |
| `author` | string | Per-example anonymised author ID starting at `u1` |
| `content` | string | Source comment body text |

Important format choices:

- The root/user prompt is stored as `input`, not as a discussion node.
- Discussion node authors start at `u1`; `u0` is reserved conceptually for the
  prompt author and does not appear in `discussion.nodes`.
- New rows omit `output` entirely.
- `metadata` is for auditing and curation; it is not intended as model input.

## Uses

This preflight slice is most appropriate for:

- smoke-testing SFT data loaders and chat templates,
- validating Parquet upload and viewer behavior,
- running small pilot training jobs,
- checking downstream consumers before the full dataset is generated.

It is not intended as:

- the final `full_v1` dataset,
- a complete assistant instruction-following dataset,
- a final-answer or QA dataset,
- a safety-filtered assistant dataset,
- a factual benchmark.

## Bias, Risks, and Limitations

The dataset inherits the risks of the public Tildes crawl, including offensive
language, misinformation, sensitive topics, copyrighted quotations, and
personally identifying information inside post content.

The source text is anonymised only in structured `author` fields. Exact source
text, URLs, topic IDs, quotations, names, or handles may still appear in content
or metadata.

DeepSeek was used only for root and branch appraisal. It did not generate,
rewrite, summarise, or trim training text. However, its decisions influence
which records appear in this slice.

Because this is a partial preflight slice, distributions by topic, group,
prompt style, and discussion shape may differ from the final full run.

## License and Redistribution

No license was identified for the source forum content. This derived dataset is
therefore marked as `license: unknown`.

The dataset consists of transformed public forum posts authored by third
parties. Users are responsible for determining whether their intended use is
lawful, including with respect to copyright, privacy, data protection, and
platform terms.