daviestechlabs/ttrpg-text-training

TTRPG Text Training

A public v0.1 research dataset from DaviesTechLabs for training a text assistant to help at a tabletop roleplaying game. The examples cover rules questions, session recaps, and what a character can reasonably know from the conversation so far.

Start with table talk

For a first supervised fine-tuning run, use reviewed_table_play_sft. Its 791 model-reviewed examples are split into 467 for training, 156 for validation, and 168 for testing. Each row has a conversation in prompt and a target reply in completion. Apply loss to the target reply. The companion reviewed_table_play_tokens configuration supplies the corresponding masked labels for the pinned Qwen chat template.

from datasets import load_dataset

train = load_dataset(
    "daviestechlabs/ttrpg-text-training",
    "reviewed_table_play_sft",
    split="train",
)
example = train[0]
print(example["prompt"][-1]["content"])
print(example["completion"][0]["content"])

If you want to train on	Configuration	Distinct examples across splits
The full supervised chat collection	table_sft_union	7,527
Preferred and rejected replies to the same question	balanced_preferences	273
Continuations grounded in public-domain book editions	gutenberg_chapter_continuations	14,918

The reviewed dialogue subset already appears in the full chat collection. The 41 configurations include alternate formats and overlapping selections. Adding their row counts would overstate the size of the dataset. The preference split contains 163 training pairs, 56 validation pairs, and 54 test pairs.

Limitations of v0.1

• Review coverage. The reviewed dialogue subset passed automated checks with the full table context. Human ratings cover only specific examples. Thirty-three contested replies from the latest review were held out of the reviewed subset. The broader configurations have different review histories.
• Preference shortcuts. Balancing answer length helps, but phrasing can still reveal the preferred reply. On an earlier 239-pair snapshot, a style-only classifier picked 36 of 50 test winners without the question. For a direct check of context use, try context_reversal_preferences, where the preferred reply changes when the session history changes.
• Rules and genre. Use the edition or house rule stated in each example. The new spell rulings use SRD 5.2.1. A ruling or an NPC’s knowledge in one campaign may change in another. Historical literary prose also differs from present-day table speech.
• Evaluation scope. The downloadable evaluation contains 44 original questions with a separate answer key. Keep those questions out of fitting and present both answer orders when scoring. These checks measure performance on a small benchmark. The release provides no estimate of improvement in a live campaign or on a newly trained model. Blind evaluation kit.

Sources and reuse

This collection combines original synthetic dialogue, openly licensed tabletop material, credited research material, and edition-specific Project Gutenberg passages. Reuse terms vary by component. Consult the credits and notices for the material you use. Project Gutenberg’s public-domain determinations are specific to the United States. Attribution and source terms.

Configuration notes

The following sections document how individual views were built. Counts described for earlier releases refer to those snapshots. The overview above gives v0.1 totals.

For grounded table dialogue, load table_completions and train on the completion only. Earlier messages are input context. Provenance records give the final answer's token interval for the pinned Qwen template. Every row fits a 1,024-token context. Use the existing train, validation and test splits.

history_completions is a subset for testing recall. Each family's two histories end with the same question and require different answers. screened_primary is the broader conversational corpus with cross-split lookalikes held from lower-priority folds. The screen combines surface overlap and answer-substitution review. It does not establish semantic independence of the entire corpus.

balanced_preferences is the preference-training view. The new pairs include rules, scene evidence and table requests. Their public source component includes stable IDs, original prompts, both answers, word counts, rationales, and section citations for rule claims. matched_preferences retains the previously verified subset with joint word and Qwen-token balance. These labels come from synthetic supervision.

source_studies contains source-linked research questions and paraphrases. The page list and associated provenance resolve every source passage used for each answer. research_notes and historical_workshops retain the prior research and workshop material.

The default table view and the all_primary archive provide message-form rows for compatibility. The views overlap. Summing them double counts examples. Counts. Attribution and individual source terms.

Source-grounded recaps

Use grounded_recaps to train or evaluate longer, multi-turn session recall. Each completion is also in table_completions. The linked provenance distinguishes the source observation from the fictional session. Each contributing work and its credited authors belong to a single split across the known source groups. The last assistant message supplies the loss target. Earlier assistant messages supply context. Source passages. Credits.

The previous table-decision release extends balanced_preferences. The winning replies also appear in table_completions in the same split. Each new preference split has equal numbers of longer and shorter chosen answers by word count. Nine-field JSONL. Source bindings.

Session evidence preferences

The additional pairs test rules, scene evidence, table constraints, and tone. Each split includes equal numbers of longer and shorter winners by word count. The raw file records the rationale, section for each rules claim, and word counts for each answer. These preference labels come from synthetic supervision. JSONL. Checks. Credits.

Research, recall, and the combined SFT view

The table_sft_union configuration contains all four final-turn training objectives described below. It merges the screened primary conversations and curated table completions. It retains one target per normalized visible dialogue, holds conflicting answers, checks rejected preferences, and reserves overlapping contexts for the evaluation fold first. Provenance records retain the original source members and the validated Qwen final-token range. The union overlaps its parent views. Mixture audit.

Use research_informed_sessions for original conversations in which changing earlier facts changes the answer to the same final request. Keep all versions of a scene in its assigned split. Each row links the research behind the exercise to a separate original player statement supporting the answer.

Use research_evidence_qa for answering from passages and citing the relevant group. Each input presents two groups from the same edition. The intended answer had to pass with the supporting evidence alone and fail with the other group alone. The standalone study questions also extend source_studies. The two research views overlap. The research articles do not establish facts about the fictional tabletop scenes. Credits and licenses.

The balanced preference view also contains the table choices batch. Its audit records word and token balance. The raw rows identify the precise error and the applicable rule section. These labels received model review.

Tabletop decisions and evaluation

The research-informed sessions include three versions of a situation with different earlier evidence. Each family stays in one fold. The published study supplies an exercise design. The fictional conversation supplies the facts for play. Each row records its paper and evidence turns. Source keys resolve through the edition and page indexes. Some locations use two column readers and record both column ranges. Article credits.

Use srd_adjudication_probes for an evaluation on SRD 5.2.1. Each exercise tests a change in resources, timing, distance or character features. Its answer key includes provenance in the split sidecar. Keep these prompts out of training. Source bindings.

Use generalization_eval to reproduce the historical benchmark on a subset of the union evaluation folds. It screens normalized answer repeats, repeated visible contexts and high-similarity context and answer pairs against earlier folds. Shared answers can be valid in both situations. The full evaluation folds are available in table_sft_union. Comparison policy.

The new preference examples give a reason for each choice. A local model reviewed each pair in both orders. The added examples have mirrored word lengths within each split, separately for rules and original scenarios. The audit reports tokenizer lengths separately.

Use class_rule_lookup to train or evaluate lookup from natural table questions. Use only the query as retrieval input. The answer and the two source keys are labels. Each retained class has two situations and stays in a single split within this configuration. A local reviewer checked source support in two orders. The source index identifies the applicable feature and a nearby alternative on the same page. Credits.

Evidence sufficiency and table preferences

The research evidence sufficiency exercises contain two versions of each question. Both use real passages from the same edition. The reviewer saw at most one item per edition in each display. One version supplies enough evidence to answer and cite a passage. The other asks for a supported partial answer and the specific missing detail. The two versions share a split. Each row links to verified source locations in its provenance record. The research questions also extend the source studies view. Research credits.

The new table preference examples preserve the word-length distribution in each split for both rulings and scenes. A local model selected the preferred answer in both orders. Raw pairs. Full preference audit. Credits.

Choose the task you want to train

Use table_play_sft for table-facing responses such as recaps, character speech, and rulings. literary_infill_sft reconstructs the exact missing wording in a credited edition. structured_answers_sft trains serialized answers such as resource calculations and event order. source_reading_sft groups reading, attribution, evidence selection, and research interpretation by the final task. Each row occurs in one of these four views and also in the full union. The mixture audit reports examples and final loss tokens by objective and split. The literary provenance files give two-reader character offsets for the target and its immediate context. Edition credits.

The gutenberg_narrative_passages configuration contains reference paragraphs with edition, chapter, XHTML member, digest, and character offsets. The new grounded_recaps examples adapt selected passages at a fictional table. Their provenance identifies the published paragraph and the original session update. Some questions deliberately supply only the printed account. All versions of a family stay in one split. The earlier generalization evaluation remains a frozen benchmark. Sources and terms.

Reading answers and table choices

Use reviewed_source_reading_sft when you want reading answers that passed two local checks against their learner-visible material and resolve to a packaged edition credit with contributors and rights. Row-to-credit index. The reviewer also saw substitute excerpts for the subset with an eligible same-edition comparison. We held rows with negative or unresolved support judgments out of this narrower view. The four broad objective views still partition the full union. Coverage and measured tokens.

The latest balanced_preferences additions passed blind review in both orders. Within each error type and split, each new pair of answer lengths has a matching reverse. The new source records retain the exact rationale and page references. Pairs · Aggregate preference audit · Attribution.

Corrected reading targets and actual table rulings

Use reviewed_source_reading_sft for source-checked answers. The current source_reading_sft has the same eligible cases after the latest repairs. Corrected targets also appear in their active training aliases. Held cases remain traceable through the review records and archived source components. The published tag proofs include physical EPUB ranges for both speaker names and quoted lines. Current reading checks · Answer changes.

Use recorded_table_rulings for session notes that distinguish the printed rule, a ruling made at the table, and an unused player option. Each scenario stays in one split. Its final request stays constant when the earlier table event changes. These examples also appear in the table training views. Rule pages · Attribution.

Use generalization_eval_corrected to score the current answers. It reapplies the published duplication and similarity screen to all current training, validation and test examples. The generalization_eval configuration preserves the previous benchmark for historical comparisons.

Research, session memory, and preference review

Five narrative research editions extend the source index. The source studies retain page references. The cited reading examples require choosing between two passages from the same paper. The research-informed sessions use original table events. Two versions of each retained session ask the same final question after different earlier events. Both answers passed model review in both presentation orders. Human votes are recorded separately when available. Each new research author group stays in one split. Edition credits.

New preference rows retain a matched reverse word-count combination within their error type and split. Related named rule subjects in earlier learner prompts constrain the split assignment. Pair rationales and quality checks.

Edition-bound chapter training and session handoffs

Use gutenberg_chapter_continuations for prose continuation training. Each row contains a contiguous context and target from one verified chapter. The pinned Qwen token IDs include a prefix masked with -100. The target contains the next paragraph. Use the supplied labels with causal language modeling. Windows do not share paragraphs. Work and contributor groups stay in a single split. Historical prose may reflect the attitudes of its period. Review the linked edition credits, headers and terms for the intended use and jurisdiction.

Use encounter_session_handoffs for dialogue. Each final answer passed two presentation orders on the relevant contrasts. Individual player and GM turns can qualify separately. The literary adaptations require the reader to distinguish the published passage from the fictional session. Only whole two-history pairs enter from that adaptation set. These rows also occur in the standard table views. Keep the development and test folds out of training. Rule and dialogue credits and Gutenberg edition credits.

Reviewed table-play supervision

For a small, reviewed table-voice run, load reviewed_table_play_tokens and train on its train split with the supplied labels. reviewed_table_play_sft provides the same rows as chat. Each record links to its previous table-play row and review method. During this release, two presentations were used to check 40 older replies sampled by source and split. The final review included the original system messages and the printed rules they supplied. Only replies with two controlled positive judgments enter this view. Newly authored replies passed the documented source, session or speaker comparisons. A model reviewed these examples. A human has not rated the new judgments. Review counts.

Weekly evidence and reading additions

The weekly_evidence_sessions rows pair two original histories that require different answers to the same table question. The weekly_preference_tokens rows contain matched branches for those prompts. The weekly_literary_comparisons rows require the response to match both the credited source passage and the separate game history. These configurations overlap the canonical SFT and preference views. Use each example once per objective. No human votes were assigned to the new replies. See row provenance and edition credits.

Weekly preference and table-delivery review

The reviewed_table_play_sft and reviewed_table_play_tokens views include older replies that passed the complete-context review. Download the current review counts. Historical notes above describe earlier versions.

The context_reversal_preferences and context_reversal_preference_tokens views use the same two replies under different histories. The correct reply changes with the history. Keep each pair of histories in its assigned split. The weekly_spell_preferences and weekly_spell_preference_tokens views contain distinct, source-checked SRD 5.2.1 rulings. All these preferences also occur in balanced_preferences. Use one representation of each example per objective.

For a manual review with one example per scenario and no repeated answer text, open the weekly JSONL. Each ruling cites its spell and page in the rationale. Attribution. Batch quality. The answer-only audit measures the preceding 239-pair corpus.

Spell components and independent evaluation

The component_ledger_preferences view combines the material-component rule with specific spell entries. The paired token view shares the prompt across both branches. The canonical balanced_preferences configuration includes these rows. See per-record sources, SRD attribution, and the complete weekly JSONL. Each scenario in the JSONL occurs once.

The separate 44-question blind evaluation stores the answer key apart from the prompts. Keep its cases out of training. The research-method index contains physical locations, and the edition credits identify the source papers.

Complete preference tensors

Load all_preference_tokens for one row per canonical pair and the original train, validation, and test splits. Each row stores the shared prompt once and two different completion suffixes. The combined sequence is prompt_input_ids plus the corresponding completion IDs. Score only suffix tokens in the preference loss.

The optional style_weighted_train_preferences view adds one mean-one sampling weight per train pair. Use a weighted loss or weighted sampling once. For a blind baseline, run score_table.py from the bundled evaluation ZIP on the held-out questions. The calibration changes how much each example counts. See the train-only weight audit and token checks.