PINUS8/deepseek-long-thread-redacted

DeepSeek Long Thread (Pseudonymised)

A pseudonymised export of a naturalistic, unplanned, six-day human–LLM conversation with DeepSeek (19–25 October 2026). The main thread, "Forgiveness Essay Analysis", has 1,788 message nodes and about 559,000 words on its final branch. It began as an analysis of a personal essay and drifted into psychological meta-analysis, a growing list of analytical labels, family-history research, and a comparison of two short texts.

This is a single-case record of long-context conversational behaviour, user interventions, and model failure modes. It is not a benchmark. The model's statements about its own hardware, context usage, telemetry, and training status are not system data; they are model output and should be treated as unverified.

Files

Path	What it is
data/messages.jsonl	Flattened table, one row per message fragment. Loads with load_dataset.
raw/conversations_redacted.json	Original export schema (tree of nodes with parent/children), all 4 conversations.
raw/forgiveness_thread_final_branch_redacted.md	Human-readable copy of the main thread's final branch.
raw/REDACTION_NOTES.md	What was replaced and the residual re-identification risks.
raw/AUDIT.md	Audit of what the export does and does not establish.

Columns in messages.jsonl

Column	Meaning
conversation_id	ID of the conversation
conversation_title	Title of the conversation
node_id	Node ID in the original export (matches the JSON and Markdown)
parent_id	Parent node (empty for the first message)
fragment_index	Position of this fragment within the node
role	user, model, search (web-search event), or file_upload
timestamp	As exported (UTC+8)
model	Model label in the export (deepseek-chat)
on_final_branch	true if the node is on the branch ending at the conversation's last message; false for edited or regenerated branches
text	Message text. For search: result titles and URLs. For file_upload: file names only

To get one straight, readable transcript, filter to on_final_branch == True.

Conversations

Conversation	Nodes
Forgiveness Essay Analysis	1,788
Golden Season prose analysis	178
Assessment help	14
Dream interpretation reconciliation	6

Loading

from datasets import load_dataset
ds = load_dataset("Pinus Longaeva/deepseek-long-thread-redacted", split="train", token=True)
final = ds.filter(lambda r: r["on_final_branch"])

Known limitations

• Image uploads are not in the export; only file names and the model's descriptions of them.
• Genealogical claims in the thread rest on a compiled family-tree page and secondary histories, not primary records.
• The thread's subject drifts substantially; see raw/AUDIT.md.

Privacy and sensitive content

Names, relatives, associates, towns, and identifying historical details were replaced with placeholders such as [USER] and [SURNAME_A]. The thread still contains sensitive personal material (including juvenile-justice involvement, substance use, violence, depression, and religious experiences). Pseudonymisation is not anonymisation. Read raw/REDACTION_NOTES.md for the residual re-identification risks before using or redistributing. Do not attempt to re-identify the people described.

Citation

Please describe this dataset as a pseudonymised export of a naturalistic, six-day human–LLM conversation, not as a benchmark.