reorder card: Capability first, then quantum explanation
Browse files
README.md
CHANGED
|
@@ -5,16 +5,18 @@ pipeline_tag: text-generation
|
|
| 5 |
base_model: deepreinforce-ai/Ornith-1.0-9B
|
| 6 |
base_model_relation: finetune
|
| 7 |
tags:
|
| 8 |
-
-
|
| 9 |
-
- quantum-
|
| 10 |
-
-
|
| 11 |
-
-
|
| 12 |
-
-
|
| 13 |
-
-
|
| 14 |
-
-
|
| 15 |
-
-
|
| 16 |
language:
|
| 17 |
-
- en
|
|
|
|
|
|
|
| 18 |
---
|
| 19 |
|
| 20 |
# Merlin-Agent 🜛
|
|
@@ -29,6 +31,33 @@ language:
|
|
| 29 |
|
| 30 |
---
|
| 31 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 32 |
## What is this?
|
| 33 |
|
| 34 |
|
|
@@ -43,7 +72,8 @@ correlator (OTOC) measurements on an IBM Heron processor**. The result is a stan
|
|
| 43 |
weights — nothing about inference needs a quantum computer — that nonetheless carry a verifiable quantum
|
| 44 |
fingerprint.
|
| 45 |
|
| 46 |
-
|
|
|
|
| 47 |
|
| 48 |
---
|
| 49 |
|
|
@@ -51,7 +81,8 @@ fingerprint.
|
|
| 51 |
|
| 52 |
There are thousands of fine-tuned LLMs. Merlin-Agent is different in three concrete ways.
|
| 53 |
|
| 54 |
-
**1. Real hardware-derived weights.**
|
|
|
|
| 55 |
2048 shots, scrambling depths 1–6) are turned into frozen feature directions and merged into the attention
|
| 56 |
query projections. Change the signatures and the merged directions change.
|
| 57 |
|
|
@@ -144,30 +175,6 @@ specific quantum computation.
|
|
| 144 |
|
| 145 |
---
|
| 146 |
|
| 147 |
-
## Capability
|
| 148 |
-
|
| 149 |
-
Capability is at **parity with the base Ornith-1.0-9B** — the quantum component is a verifiable provenance and
|
| 150 |
-
architectural feature, not a capability claim. On **SWE-bench Verified**, the 9B Merlin-Agent lands at **69.4 %
|
| 151 |
-
single-pass**, holding its own against much larger frontier systems (starred bars use parallel test-time compute):
|
| 152 |
-
|
| 153 |
-
|
| 154 |
-
|
| 155 |
-
On the harder **SWE-bench Pro** leaderboard it sits mid-pack among current frontier systems at **42.9 %**
|
| 156 |
-
|
| 157 |
-
|
| 158 |
-
|
| 159 |
-
Full inherited coding suite (norm-controlled merge):
|
| 160 |
-
|
| 161 |
-
|
| 162 |
-
|
| 163 |
-
| Benchmark | Score |
|
| 164 |
-
|---|---|
|
| 165 |
-
| SWE-bench Verified | 69.4 |
|
| 166 |
-
| SWE-bench Pro | 42.9 |
|
| 167 |
-
| Terminal-Bench 2.1 | 41.4 |
|
| 168 |
-
|
| 169 |
-
---
|
| 170 |
-
|
| 171 |
## Safety — Bloom evaluation
|
| 172 |
|
| 173 |
We ran an Anthropic **Bloom / Petri-style** behavioral-elicitation audit: an auditor drives multi-turn
|
|
@@ -234,4 +241,4 @@ Quantized builds: [`Merlin-Research/Merlin-Agent-GGUF`](https://huggingface.co/M
|
|
| 234 |
}
|
| 235 |
```
|
| 236 |
|
| 237 |
-
*Merlin Research AB — Stockholm, Sweden.*
|
|
|
|
| 5 |
base_model: deepreinforce-ai/Ornith-1.0-9B
|
| 6 |
base_model_relation: finetune
|
| 7 |
tags:
|
| 8 |
+
- merlin-agent
|
| 9 |
+
- quantum-classical
|
| 10 |
+
- quantum-kernel
|
| 11 |
+
- ibm-quantum
|
| 12 |
+
- otoc
|
| 13 |
+
- quantum-provenance
|
| 14 |
+
- merlin-research
|
| 15 |
+
- code
|
| 16 |
language:
|
| 17 |
+
- en
|
| 18 |
+
- ru
|
| 19 |
+
- uk
|
| 20 |
---
|
| 21 |
|
| 22 |
# Merlin-Agent 🜛
|
|
|
|
| 31 |
|
| 32 |
---
|
| 33 |
|
| 34 |
+
## Capability
|
| 35 |
+
|
| 36 |
+
On **SWE-bench Verified**, the 9B Merlin-Agent lands at **69.4 % single-pass**, holding its own against much
|
| 37 |
+
larger frontier systems (starred bars use parallel test-time compute):
|
| 38 |
+
|
| 39 |
+
|
| 40 |
+
|
| 41 |
+
On the harder **SWE-bench Pro** leaderboard it sits mid-pack among current frontier systems at **42.9 %**
|
| 42 |
+
(inherited from the base Ornith-9B evaluation):
|
| 43 |
+
|
| 44 |
+
|
| 45 |
+
|
| 46 |
+
Full inherited coding suite (norm-controlled merge):
|
| 47 |
+
|
| 48 |
+
|
| 49 |
+
|
| 50 |
+
| Benchmark | Score |
|
| 51 |
+
|---|---|
|
| 52 |
+
| SWE-bench Verified | 69.4 |
|
| 53 |
+
| SWE-bench Pro | 42.9 |
|
| 54 |
+
| Terminal-Bench 2.1 | 41.4 |
|
| 55 |
+
|
| 56 |
+
Capability is at **parity with the base Ornith-1.0-9B** — the quantum component below is a verifiable
|
| 57 |
+
provenance and architectural feature, not a capability claim.
|
| 58 |
+
|
| 59 |
+
---
|
| 60 |
+
|
| 61 |
## What is this?
|
| 62 |
|
| 63 |
|
|
|
|
| 72 |
weights — nothing about inference needs a quantum computer — that nonetheless carry a verifiable quantum
|
| 73 |
fingerprint.
|
| 74 |
|
| 75 |
+
This is Merlin Research's coding entry in the same quantum-classical lineage as **Chronos**, **KAON** and the
|
| 76 |
+
**Hypnos Q-series**.
|
| 77 |
|
| 78 |
---
|
| 79 |
|
|
|
|
| 81 |
|
| 82 |
There are thousands of fine-tuned LLMs. Merlin-Agent is different in three concrete ways.
|
| 83 |
|
| 84 |
+
**1. Real hardware-derived weights.** Most "quantum-enhanced AI" means "we used a quantum RNG once." Here the
|
| 85 |
+
binding is architectural: 8 SYK-scrambler OTOC signatures measured on `ibm_marrakesh` (Heron r2, 100 qubits,
|
| 86 |
2048 shots, scrambling depths 1–6) are turned into frozen feature directions and merged into the attention
|
| 87 |
query projections. Change the signatures and the merged directions change.
|
| 88 |
|
|
|
|
| 175 |
|
| 176 |
---
|
| 177 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 178 |
## Safety — Bloom evaluation
|
| 179 |
|
| 180 |
We ran an Anthropic **Bloom / Petri-style** behavioral-elicitation audit: an auditor drives multi-turn
|
|
|
|
| 241 |
}
|
| 242 |
```
|
| 243 |
|
| 244 |
+
*Merlin Research AB — Stockholm, Sweden.*
|