schonsense/human_ai_story_contrastive_v2

Collapsed and Augmented with llama3.3sft on-policy/policy adjacent responses

Prepare for some AI generated descriptions.

Dataset Construction and Statistics

• Purpose: A contrastive creative-writing dataset for studying distributional stylistic differences between human-written and model-generated text.

• Total size: 5,733 texts grouped across 1,443 unique prompts / human responses.

• Source breakdown:

  • 1,443 human responses
  • 1,421 GPT-3.5 responses
  • 1,426 Claude Opus responses
  • 1,443 starting-policy (pi0) responses

• Grouped structure: Responses associated with the same original writing prompt share a common group_id, allowing prompt-disjoint train/validation/test splits and preventing alternate responses to the same prompt from leaking across evaluation partitions.

• Human response deduplication: The original paired dataset contained repeated human stories where the same human response had been paired independently with GPT-3.5 and Opus outputs. These were collapsed into a single grouped record keyed by the normalized prompt + human response.

• Machine sources remain distinct: GPT-3.5, Opus, and pi0 are preserved as separate response populations rather than being merged into one generic machine class.

• Starting-policy augmentation: One additional response was generated for every unique prompt using the unmodified target model, producing a complete 1,443-response pi0 population.

• No manual cleanup of pi0 outputs: Starting-policy generations are retained as produced so that they reflect the model's native creative-writing distribution.

• Sequence-length statistics under the target model tokenizer/chat template:

  • Minimum full sequence: 124 tokens
  • Median full sequence: 933 tokens
  • Maximum full sequence: 3,688 tokens
  • Minimum response: 52 tokens
  • Median response: 848 tokens
  • Maximum response: 3,580 tokens
  • Total tokenized corpus: 5,750,556 tokens

• Style representation: Complete responses were embedded using StyleDistance, with long responses represented through chunked embeddings pooled across the full text.

• Human/machine separability: Human and machine responses are strongly distinguishable in StyleDistance space, while response length alone does not explain the distinction.

• Cross-model structure: Human-vs-machine style boundaries learned from some machine sources transfer strongly to unseen machine families, indicating that the dataset contains both model-specific stylistic structure and a broader shared machine-writing component.

• Important methodological finding: Apparent prompt-local correction vectors between machine and paired human responses did not survive anchor-preserving null tests. The dataset is therefore best treated as representing population-level stylistic distributions, rather than as a collection of exact machine→human rewrite directions.

• Intended use: Suitable for research on:

  • human vs. machine creative-writing style;
  • cross-model stylistic transfer;
  • model-family-specific vs. shared machine-writing characteristics;
  • style embedding and representation analysis;
  • distributional comparison of human and generated writing.

Dataset / Methodology Highlights

• Core structure: The dataset contains 1,443 unique writing prompts / human responses, augmented with machine-written responses to the same prompts:

  • 1,443 human responses
  • 1,421 GPT-3.5 responses
  • 1,426 Claude Opus responses
  • 1,443 responses from the target starting policy (pi0)
  • 5,733 total texts

• Human responses are not treated as exact imitation targets. The objective is not to teach the model to reproduce the particular human continuation associated with a prompt. Human writing instead defines a broader stylistic distribution against which machine-output distributions can be compared.

• Multiple machine populations are intentionally retained. GPT-3.5, Opus, and pi0 are kept as distinct sources rather than pooled during data construction. This makes it possible to distinguish:

  • model-family-specific stylistic artifacts;
  • stylistic properties shared across machine-generated writing;
  • characteristics specific to the target starting policy.

• External style representation: Each complete response was embedded using StyleDistance, a content-resistant style embedding model designed to capture stylistic rather than primarily semantic similarity.

• Full-response analysis: Long responses are represented using chunked StyleDistance embeddings pooled across the complete text. The style target therefore represents the entire response rather than only its opening or a fixed token window.

• Strong human/machine separability: Human and machine texts are highly distinguishable in StyleDistance space, while simple response length does not explain the separation. Cross-model tests also show that a human-vs-machine boundary learned from some model families transfers strongly to previously unseen machine families.

• The human/machine distinction is not reducible to a single centroid direction. Human writing occupies a broader and higher-dimensional stylistic region, while different machine families occupy distinct but partially shared regions. Consequently, the methodology does not simply attract outputs toward a human centroid or repel them from a machine centroid.

• Prompt-local vector arithmetic was explicitly tested and rejected. Initial apparent alignment between (human − GPT) and (human − Opus) vectors disappeared under anchor-preserving null tests. The resulting methodology therefore treats the human/machine distinction as a distributional field, not as a per-prompt correction vector pointing toward the paired human response.

• Cross-model transfer supports a shared machine-style component. A StyleDistance classifier trained on GPT-3.5 + Opus generalizes strongly to unseen pi0 responses, indicating that the dataset contains a stylistic signal shared across substantially different model families rather than only model-specific fingerprints.

• Scalar style field: A continuous human-style score is derived from the StyleDistance human-vs-machine boundary. Higher values correspond to regions of style space that are more characteristic of human writing.

• Hidden-state surrogate: To make this external style field usable during model training, the target model's teacher-forced hidden trajectories were analyzed. For each response, the final normalized hidden states that causally predict the assistant response tokens are pooled into a single mean hidden-state representation.

• A simple linear surrogate proved sufficient. The mean final predictive hidden state contains a strong linearly accessible approximation of the StyleDistance field. Adding hidden-state standard deviation did not improve performance, suggesting that the useful signal is primarily encoded as a direction in the mean final predictive representation.

• Machine-local critic: The final surrogate is trained on GPT-3.5 + Opus + pi0 hidden trajectories, while subtracting each machine source's mean style score before regression. This prevents easy between-model offsets from dominating the critic and forces it to learn stylistic variation shared within machine populations.

• Human hidden states are not required for the final critic. Human examples remain essential for defining the external StyleDistance field, but including human hidden-state trajectories in the surrogate regression reduced fidelity within the target pi0 population.

• Strong target-policy validation: Using 5-fold group_id-disjoint cross-validation, the final source-centered machine critic reaches approximately:

  • π0 Spearman ρ ≈ 0.619
  • GPT-3.5 Spearman ρ ≈ 0.587
  • Opus Spearman ρ ≈ 0.643

  This indicates that the learned direction captures continuous stylistic variation shared across machine families rather than merely identifying which model produced a response.

• Regularization is important. Strongly regularized ridge regression substantially outperforms weakly regularized fits. The selected critic uses a ridge strength of 10,000, chosen by group-disjoint cross-validation to maximize held-out pi0 rank correlation.

• Training signal is reference-relative rather than an unbounded reward. The intended use is to compare the current model's hidden trajectory with the cached starting-policy trajectory for the same response and encourage a modest movement along the validated human-style direction.

• Locality is deliberate. The critic is validated around the starting-policy distribution and should not be interpreted as a globally reliable measure of writing quality. Training targets are therefore bounded to remain within the empirically observed pi0 score distribution rather than maximizing the critic indefinitely.

• Current calibrated intervention scale: The natural standard deviation of critic scores within pi0 is approximately 0.187. A first intervention target of roughly +0.094 critic units corresponds to a half-standard-deviation movement while remaining within the densely populated region of the starting-policy distribution.

• No response-level NLL is required by the methodology. The intended intervention applies distributional pressure toward a more human-like region of stylistic space without requiring the model to increase the likelihood of the specific paired human continuation.

• Primary research use: The dataset is intended for experiments in:

  • human-vs-machine writing-style representation;
  • cross-model style-transfer analysis;
  • hidden-state style probes and critics;
  • distributional fine-tuning objectives;
  • preference/style optimization that avoids exact continuation imitation;
  • analysis of shared vs model-specific stylistic artifacts.

• Important interpretation: A high critic score should be understood as movement toward the human-associated region learned from this dataset, not as a universal measure of literary quality, correctness, creativity, or human authorship.