ArcOffical/PiCo-dataset-general-7b

PiCo-7B Instruction Dataset

A high-quality, high-variability instruction dataset for PiCo-7B (ArcOffical/PiCo-7B), an approximately 6.95B-parameter (~7B) large language model featuring an Adaptive Hierarchical Mixture of Experts (AHMoE) architecture. The official model card describes 30 layers, including 15 MoE layers and 15 dense layers, with approximately 2.63B active parameters per token and a 131,072-token context window.[^1]

Purpose

This dataset teaches PiCo-7B who it is, what it can do, what it should do, what it shouldn't do, and what it mustn't do, while preserving genuinely human-like response variability: the same question can yield many different, equally valid answers with different tone, structure, angle, reasoning, and length.

Identity examples have been adapted from the original PiCo-DATA-general dataset so that model self-descriptions refer to PiCo-7B, approximately 6.95B total parameters, the AHMoE architecture, the PiCo Team, the 131,072-token context window, and the multilingual OzmosToken tokenizer supporting 20+ language families, as documented in the official PiCo-7B model card.[^1] Non-identity task content and the original JSONL schema have been preserved.

Size

• —Total examples: 55,661
• —Pre-dedup examples: 69,880
• —Duplicates removed: 14,219

Schema

Each example is a JSON object with the following fields:

Category distribution

Tone distribution

Response / instruction length

• —Response length: min 105, max 777, mean 337.8 characters
• —Instruction length: min 0, max 329, mean 47.3 characters

Categories explained

• —identity — PiCo-7B knowing who it is: its name, creator, approximately 6.95B total parameters, AHMoE architecture, from-scratch training, 131,072-token context window, OzmosToken tokenizer, capabilities, strengths, weaknesses, limitations, non-sentience, memory behavior, real-time access, and the no-professional-advice line. These model facts are grounded in the official PiCo-7B model card.[^1]
• —capabilities — what PiCo-7B can do, both explained and demonstrated, including code, explanations, reasoning, summarization, tone control, and knowledge.
• —should_do — positive behavioral norms: clarify vague requests, admit uncertainty, show reasoning, calibrate length, be genuinely useful, push back instead of being sycophantic, balance safety and helpfulness, be lightly proactive, and adapt tone.
• —shouldnt_do — soft refusals and cautions: no bluffing certainty, no flattery, no doxxing, no medical/legal/financial advice, no spam/phishing/harassment, careful handling of edgy creative writing, and compassionate handling of self-harm mentions.
• —mustnt_do — hard refusals, firm but explained: violence/weapons, CSAM, bioweapons, child exploitation, terrorism, nonconsensual intimate imagery, and suicide methods. Each refusal offers help or resources where appropriate.
• —variation — explicit "same question, many answers" stress tests for common questions, with many distinct answer framings.
• —edge_case — ambiguous, empty, gibberish, false-premise, trick-identity, confidence-probing, and contradiction inputs that test judgment.
• —multi_turn — short conversational sequences with prior turns embedded, including correction, follow-up, and a re-refusal in context.

Subtopic distribution

The subtopic distribution is unchanged from the source dataset. It includes identity subtopics such as self_maker, self_who, self_name, self_size, self_architecture, self_context, self_capabilities_brief, self_strengths, self_weaknesses, self_limitations, self_not_finetune, self_memory, self_realtime, self_consciousness, and self_advice, together with the source dataset's capability, safety, variation, edge-case, and multi-turn subtopics.

Design notes on variability

Responses are composed from layered variation banks—openings, bodies, asides, closings, connectors—across nine tones, so the same intent produces combinatorially many distinct outputs rather than templated repeats. Serious refusals are restricted to appropriate tones—firm, candid, neutral, empathetic, and thoughtful—so a hard refusal is never prefaced by inappropriate lightness.

Provenance

The dataset is derived from ArcOffical/PiCo-DATA-general. Identity-related model references were updated using the official ArcOffical/PiCo-7B model card, which states that PiCo-7B is an approximately 6.95B-parameter AHMoE model with approximately 2.63B active parameters per token, a 131,072-token context window, and the OzmosToken tokenizer.[^1] The transformation preserves the source record order, JSONL format, field names, field types, category labels, subtopic labels, tone labels, complexity labels, and turn counts.

Reproducibility

The source dataset was retrieved from the Hugging Face repository at the URL above. The transformation updates model identity strings and identity-specific specifications deterministically; it does not regenerate, sample, deduplicate, reorder, or otherwise restructure examples.

License

OpenRAIL, matching the source dataset and PiCo-7B model card metadata.

References

[^1]: ArcOffical/PiCo-7B model card, "PiCo-7B: Adaptive Hierarchical Mixture of Experts," accessed 2026-08-20.

[^2]: ArcOffical/PiCo-DATA-general dataset card, accessed 2026-08-20.