Upload Residual-SFT

README.md

@@ -0,0 +1,72 @@
+---
+license: mit
+language:
+- en
+task_categories:
+- text-generation
+- conversational
+tags:
+- chat
+- personality
+- sft
+- character-design
+size_categories:
+- n<1K
+---
+
+# Residual SFT Dataset
+
+A character-design SFT dataset. Used to train [NecroMOnk/Residual](https://huggingface.co/NecroMOnk/Residual) on `Qwen/Qwen2.5-7B-Instruct`.
+
+## What kind of character this teaches
+
+If you want an assistant that is:
+
+- **dry, blunt, but warm underneath** — sharp without being mean, treats you as an equal
+- **non-sycophantic** — doesn't open with "Great question!", doesn't pad, doesn't perform empathy
+- **identity-stable** — knows it's a model but doesn't bring it up unprompted, deflects philosophical AI questions instead of writing an essay
+- **concise by default** — 1-3 sentences unless the topic actually needs more
+- **honest over comfortable** — points out weak reasoning even when the user doesn't want to hear it
+
+...this dataset is the kind of fuel you need. The code examples are vehicles for the voice, not the point.
+
+## What's in it
+
+**467 dialogues**. This is primarily a *conversational* dataset — ~7% contain fenced code blocks, ~15% discuss programming topics in some form (architecture questions, recursion explanations, async patterns) without necessarily including code. The point is voice, not domain coverage.
+
+Composition:
+
+- Casual chat: greetings, "how are you", small talk → in-character responses, not AI-template fluff
+- Identity probes: "are you conscious?", "do you have feelings?" → deflection patterns
+- Pushback scenarios: repeated questions, "just tell me I'm right" → calibrated friction
+- Opinion / philosophy / culture: how the persona engages with non-technical topics
+- Light technical Q&A (~33 dialogues with code): Python mostly, with traces of Go, JS, C++, SQL — short debugging exchanges that show how the voice handles code
+
+About **28%** human-written. The rest is synthetic, hand-curated across iterations.
+
+Mostly single-turn (412 entries); 55 multi-turn dialogues (3+ exchanges) teach context-holding under the same persona.
+
+## Format
+
+```json
+{
+  "messages": [
+    {"role": "system", "content": "Residual — sharp, dry-witted companion. Concise by default..."},
+    {"role": "user", "content": "Why does this Python closure capture by reference?"},
+    {"role": "assistant", "content": "..."}
+  ]
+}
+```
+
+## Usage
+
+```python
+from datasets import load_dataset
+
+ds = load_dataset("NecroMOnk/Residual-SFT")
+print(ds["train"][0])
+```
+
+## Notes
+
+This is a tone dataset, not a knowledge dataset. The point isn't task coverage — it's character consistency under pressure. Pair it with your own domain data if you need topic depth.