georgep123 commited on
Commit
eaa4aa1
·
verified ·
1 Parent(s): 2fc6da5

Add 8B/30B family comparison and which-model-to-use guidance; state plainly that the head-to-head capability comparison is not yet drawable

Browse files
Files changed (1) hide show
  1. README.md +37 -0
README.md CHANGED
@@ -117,6 +117,43 @@ an edit that resolves the finding *and* leaves the file valid — which is why t
117
  model gets each one right is exactly what the evaluations listed below will establish, and those
118
  numbers are not published here yet.
119
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
120
  ## At a glance
121
 
122
  | Property | Value |
 
117
  model gets each one right is exactly what the evaluations listed below will establish, and those
118
  numbers are not published here yet.
119
 
120
+ ## Which model should you use
121
+
122
+ There are two models in this family. They were trained on **the same 96-row corpus with the same
123
+ recipe** — same rank, same scaling, same loss, same three behaviours. The only deliberate difference
124
+ is the size of the parent they were applied to.
125
+
126
+ | | [Vinci Cyber 8B 1.0](https://huggingface.co/simpledirect/Vinci-Cyber-8B-1.0) | **Vinci Cyber 30B 1.0** (this model) |
127
+ |---|---|---|
128
+ | Parameters | ~8B | ~28.87B |
129
+ | Parent | Granite 4.1 8B | Granite 4.1 30B |
130
+ | Layers | 40 | 64 |
131
+ | Weights on disk | ~16.8 GB, one file | ~57.7 GB, two shards |
132
+ | Memory in bf16 | roughly 17 GB | roughly 58 GB before runtime overhead |
133
+ | Runs on | a single 24 GB card | a single 80 GB card, or two smaller ones |
134
+ | Training corpus | identical — 96 rows, same digest | identical — 96 rows, same digest |
135
+ | Training recipe | identical — DoRA + rsLoRA, r=32, α=64 | identical — DoRA + rsLoRA, r=32, α=64 |
136
+ | IaC repair result | published: parent 0/8 → **4/8**, restraint 22/22 | **measured, publication held** pending scorer qualification |
137
+ | Licence | Apache-2.0 | Apache-2.0 |
138
+
139
+ **Use the 8B if** you want the published, reviewed result today; you are running on one consumer or
140
+ mid-range card; you are putting the model in a loop where latency matters; or you are evaluating
141
+ whether this approach fits your workflow at all. It is the model with numbers you can check right now.
142
+
143
+ **Use the 30B if** you have the memory budget and want the larger parent's headroom — more general
144
+ knowledge behind the same specialization, and more capacity for longer files and more context around
145
+ a finding.
146
+
147
+ 🔴 **What we will not tell you yet: which one repairs more findings.** The 30B's in-domain result is
148
+ measured and held back pending independent qualification of the scoring instrument, so a head-to-head
149
+ capability comparison cannot honestly be drawn here. Bigger parent is a reason to *expect* more; it
150
+ is not evidence of it, and we are not going to dress expectation up as a result. When the instrument
151
+ is qualified, this section gets the comparison — or it gets the news that scale did not help.
152
+
153
+ If you need broad CVE repair across languages rather than IaC security findings, **use something
154
+ else.** The 8B's out-of-domain benchmark showed no measurable improvement over its parent, and
155
+ nothing suggests the 30B changes that. This family is narrow on purpose.
156
+
157
  ## At a glance
158
 
159
  | Property | Value |