Update README.md
Browse files
README.md
CHANGED
|
@@ -1,3 +1,41 @@
|
|
| 1 |
-
---
|
| 2 |
-
|
| 3 |
-
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
language:
|
| 3 |
+
- en
|
| 4 |
+
license: mit
|
| 5 |
+
pretty_name: Teleportation & Wormholes — Evidence-Only Scientist Mindframe
|
| 6 |
+
tags:
|
| 7 |
+
- physics
|
| 8 |
+
- general-relativity
|
| 9 |
+
- quantum-teleportation
|
| 10 |
+
- rag
|
| 11 |
+
- dpo
|
| 12 |
+
- chain-of-thought
|
| 13 |
+
---
|
| 14 |
+
|
| 15 |
+
# Teleportation & Wormholes — Evidence-Only Scientist Mindframe (Within Us AI)
|
| 16 |
+
|
| 17 |
+
## What this is
|
| 18 |
+
A multi-dataset package built from **your PDFs only**. It trains:
|
| 19 |
+
- Evidence-only scientific answering (no made-up facts)
|
| 20 |
+
- Deep reasoning with falsification and verification plans
|
| 21 |
+
- Tool-first retrieval behavior (RAG-Bridge)
|
| 22 |
+
- Preference shaping against sci-fi/hallucination patterns (DPO)
|
| 23 |
+
- Agentic, multi-expert thinking (MoE-style contributions) without role-play fluff
|
| 24 |
+
|
| 25 |
+
## Critical rule
|
| 26 |
+
**Facts must be directly supported by the provided chunk evidence.**
|
| 27 |
+
When proposing "inventions" or "discoveries", outputs MUST label them as:
|
| 28 |
+
- Hypothesis / Proposal
|
| 29 |
+
- Constraints from evidence
|
| 30 |
+
- Validation plan (what would confirm/refute)
|
| 31 |
+
|
| 32 |
+
## Datasets
|
| 33 |
+
- `data/sft` (50k)
|
| 34 |
+
- `data/cot` (100k)
|
| 35 |
+
- `data/dpo` (500k)
|
| 36 |
+
- `data/rag_bridge` (1M)
|
| 37 |
+
- `data/hybrid_moe_agentic` (10M)
|
| 38 |
+
|
| 39 |
+
## Input chunk format
|
| 40 |
+
A JSONL file with lines: `{id, source, domain, page_start, page_end, text}`
|
| 41 |
+
Use the provided `chunks_clean_all.jsonl` or your own.
|