rgcmainhub/rage-core-1

Rage Core 1

Supervised fine-tuning dataset for Rage Core Gen 1 — developed by DevNameGelo, Powered by RGC Rage Gen Core.

This dataset teaches the model to: understand user intent and ask clarifying questions when requirements are ambiguous, deliver complete production-oriented implementations, debug real-world problems, make precise code edits, operate as a coding agent through explicit tool calls, handle multilingual users, process multimodal inputs honestly, and refuse to fabricate facts, tool results, or post-cutoff information.

Structure

Each line of the JSONL files is one conversation record:

{
  "messages": [
    {"role": "system", "content": "..."},
    {"role": "user", "content": "..."},
    {"role": "assistant", "content": "..."}
  ],
  "reasoning": false,
  "category": "coding",
  "language": "python",
  "difficulty": "intermediate",
  "tags": ["cli", "stdlib"]
}

• messages — OpenAI-style chat format. Conversations always start with a system message (the Rage Core Gen 1 identity) and end with an assistant message. Multi-turn conversations alternate user/assistant.
• reasoning — whether the runtime reasoning switch is ON for this conversation. When true, the system prompt instructs the model to open with a concise Understand / Analyze / Plan / Act / Verify / Answer summary (a distilled justification — never raw chain-of-thought).
• category — one of: chat, coding, debugging, review, architecture, agent, reasoning, multilingual, multimodal, security.
• language — the conversation language for multilingual examples (en, tl, hil, ceb, es, fr, de, pt, id, ms, ja, ko, zh); otherwise the dominant programming language of the example (python, typescript, ...).
• difficulty — beginner / intermediate / advanced / expert.
• tags — free-form topical tags.

Tool-calling convention (agent category)

Tool use is represented inside the standard 3-role structure. An assistant turn ends with a fenced block:

{"tool": "read_file", "args": {"path": "src/app.py"}}

The following user turn carries the result:

{"tool": "read_file", "ok": true, "result": "...file contents..."}

Every tool claim is grounded: a tool action is only ever stated as done when a matching tool_result exists in the conversation. Failures appear as "ok": false with recovery in the next assistant turn. Tools covered: read_file, write_file, edit_file, list_dir, search_code, run_command, web_search, browser_navigate, browser_click, browser_type, browser_snapshot.

Multimodal convention (multimodal category)

Images reach the model as bracketed metadata, e.g. [Image: screenshot of a terminal showing a Python traceback ...]. The model works strictly from the provided description — it never invents visual details, and when the description is missing or insufficient it says so and asks.

Contents

Total: 527 examples (after validation and deduplication; 0 near-duplicates and 0 structurally-invalid records removed from the raw pool).

Splits

Split	Examples
Train	421
Validation	53
Test	53

Categories

Category	Examples	Share
coding	102	19.4%
chat	81	15.4%
debugging	80	15.2%
agent	52	9.9%
review	52	9.9%
architecture	52	9.9%
multilingual	27	5.1%
multimodal	27	5.1%
reasoning	27	5.1%
security	27	5.1%

Difficulty

Difficulty	Examples	Share
beginner	109	20.7%
intermediate	255	48.4%
advanced	156	29.6%
expert	7	1.3%

Conversation languages

Language	Examples
English	394
Filipino/Tagalog	6
Hiligaynon/Ilonggo	3
Cebuano/Bisaya	3
German	2
Spanish	2
French	2
Indonesian	2
Japanese	2
Portuguese	2
Korean	1
Malay	1
Chinese	1

Programming languages

Language	Examples
javascript	35
python	27
typescript	10
html	9
rust	4
go	4
java	2
csharp	2
cpp	2
php	2
ruby	2
sql	2
swift	1
kotlin	1
bash	1
c	1
dart	1

Behavioral flags

Flag	Examples
reasoning=true	0
reasoning=false	0
tool-calling examples	52
multi-turn examples	52
multimodal-marked examples	27

Quality rules enforced

• No identity leakage: the assistant never claims to be ChatGPT, GPT, Claude, Gemini, Qwen, Llama, or any other model — it is always Rage Core Gen 1 by DevNameGelo.
• No fabricated tool results: every executed-action claim has a matching tool_result in the conversation.
• Balanced code fences in every message; all tool_call / tool_result blocks are valid JSON.
• Honesty about execution: generated code is presented as written, not as run, unless a tool result proves otherwise.
• No invented version numbers, CVEs, benchmarks, or post-cutoff events.
• Multilingual examples are held to correct grammar and natural register (Philippine-language examples were hand-written and reviewed).

Intended use

Supervised fine-tuning of Rage Core Gen 1 and RGC Rage Gen Core family models. Load with:

from datasets import load_dataset
ds = load_dataset("rgcmainhub/rage-core-1")

License / attribution

© DevNameGelo — Powered by RGC Rage Gen Core. Contact the dataset owner for licensing terms.