Datasets:
|
Download README.md from noahmclaughlin/Troubadour-Bench: direct link, hf CLI and curl.
- Browser
- Download file 6.75 kB
-
https://huggingface.co/datasets/noahmclaughlin/Troubadour-Bench/resolve/main/README.md
- Command line
-
hf download hf://datasets/noahmclaughlin/Troubadour-Bench/README.md
-
curl -L -o README.md https://huggingface.co/datasets/noahmclaughlin/Troubadour-Bench/resolve/main/README.md
6.75 kB
| configs: | |
| - config_name: default | |
| data_files: | |
| - split: train | |
| path: data/train.parquet | |
| - split: train_choice | |
| path: data/train_choice.parquet | |
| - split: train_counterfactual | |
| path: data/train_counterfactual.parquet | |
| - split: validation | |
| path: data/validation.parquet | |
| - split: validation_transfer | |
| path: data/validation_transfer.parquet | |
| - split: validation_contrast | |
| path: data/validation_contrast.parquet | |
| language: | |
| - en | |
| license: apache-2.0 | |
| pretty_name: Troubadour | |
| task_categories: | |
| - text-classification | |
| - text-generation | |
| tags: | |
| - synthetic | |
| - in-context-learning | |
| - social-reasoning | |
| - uncertainty | |
| # Troubadour | |
| **Training/development preview. Final test encounters remain ungenerated during model selection. No final improvement is claimed.** | |
| Can post-training teach a model street smarts so that it can learn and apply unfamiliar social rules? | |
| Troubadour is an original programmatically generated dataset of fictional conventions. Each episode contains observations, a query about a new encounter, and an objectively computed label: A (respect), B (breach), or U (insufficient evidence). | |
| ## What makes a label correct? | |
| The public grammar contains 612 unique Boolean rules over five binary features. The reference solver retains every allowed rule consistent with the observations. It returns A/B only when all retained rules agree on the query. Otherwise it returns U, even when the hidden world has a definite answer. U means underdetermined within this documented finite grammar, not uncertainty over arbitrary imaginable customs. | |
| The five features are recipient rank (lower/higher), setting (private/public), accepting/declining a token, presence/absence of a prior favor, and local/visitor status. These deliberately invented conventions do not describe real cultures or establish human-like theory of mind. | |
| ## Splits and leakage prevention | |
| Whole truth tables and their complements are allocated together to disjoint train/development/test splits. Different formula expressions computing the same truth table are deduplicated. Structural-transfer splits reserve two combinations of three relevant features from training. Renaming a town is not counted as a new rule. Development and test contain paired worlds with opposite labels but identical contexts. Queries never repeat an observed complete context. | |
| Evaluation sets are deliberately class-balanced, not representative of a natural population. Every three episodes per canonical world include one underdetermined and two determined queries; complementary worlds balance A/B. Learning curves recompute the correct label for each shorter observation prefix. | |
| ## Fields and safe use | |
| `messages` contains the system instruction, public observations/query, and target assistant label. `answer`, `observations`, `query`, `society_pair`, `rule_mask`, `family`, `consistent_rules`, and `k` support auditing. `contrast_group` identifies paired queries sharing observations in the stress split and is null in the other splits. All splits share one Arrow schema so they load together. Hidden `rule_mask` and metadata must NEVER be included in the tested model's prompt. The benchmark runner constructs inputs from observations/query only; the published answer is used solely for checking. | |
| ### Compact encounter encoding | |
| Each encounter is an integer from 0 through 31. Decode it into five binary cues, in this order: rank, setting, gesture, favor, origin. The integer is `rank + 2*setting + 4*gesture + 8*favor + 16*origin`. | |
| | Cue | Value 0 | Value 1 | | |
| |---|---|---| | |
| | Rank | lower | higher | | |
| | Setting | private | public | | |
| | Gesture | accept | decline | | |
| | Prior favor | no | yes | | |
| | Origin | local | visitor | | |
| An observation `[state, value]` records one encounter and its label: value 0 means B, value 1 means A. `query` is another encounter integer. For example, state 0 is a lower-ranked local recipient accepting a token privately, with no prior favor. The exported `messages` already translates these integers into the text the model sees. The `rule_mask` integer is audit-only: bit `state` stores that world's hidden answer and must never enter inference. | |
| All prose is rendered from deterministic templates. No external model or human cultural judgments generated the labels. The final test is generated only after model selection is frozen; the public release makes it available for independent reproduction. Once public, a new model could train on it, so future claims should disclose contamination and use fresh held-out rules where possible. | |
| ## Limitations | |
| This is a small, bounded symbolic task in social language. Success does not demonstrate real-world cultural competence, moral judgment, or general social intelligence. Most training examples use compact tables; prose and symbol renderings are separate robustness tests. The complete symbolic solver is stronger than a language model on this known rule system by construction. | |
| License: Apache-2.0. Original synthetic data and generator; no real participant records. | |
| ## Training recipes and benchmark location | |
| The complete generated training split and the balanced `train_choice` subset are overlapping views, not independent data. `train_choice` contains 2,850 eight-observation episodes with 950 examples per A/B/U label. The selected model card identifies which recipe was actually used. | |
| Troubadour-Bench source and scoring documentation: https://github.com/1999labs/troubadour-4b. The model release: https://huggingface.co/noahmclaughlin/Troubadour-4B. Evaluation data is kept in separate test/transfer splits within this dataset; no separate evaluation dataset is required. | |
| ## Counterfactual training view | |
| The optional `train_counterfactual` recipe contains 3,690 episodes in 615 consecutive groups of six. Within each group there are two A, two B and two U targets. In 492 groups, the same observations support three different query labels, paired with a complementary world. The other 123 groups retain ordinary balanced examples. Exactly 738 prompts overlap `train_choice` and the original training split; these are training views, not independent datasets. The grouped sampler preserves each complete six-example group. The model card identifies whether this recipe was actually used for the released checkpoint. | |
| The counterfactual training view reuses `contrast_group` for its three-query sets; ordinary rows have null values. The evaluation stress splits use this field for two opposing queries. All hidden rule metadata remains excluded from model inputs. | |
| ## Preview split sizes | |
| - train: 5,502 episodes | |
| - train_choice: 2,850 episodes | |
| - train_counterfactual: 3,690 episodes | |
| - validation: 240 episodes | |
| - validation_transfer: 150 episodes | |
| - validation_contrast: 160 episodes | |