add Qwen3.5 equal-bytes figure
Browse files
README.md
CHANGED
|
@@ -56,9 +56,12 @@ generative use scores lower across the board — see the
|
|
| 56 |
[retrieval results](https://github.com/OpenBallast/ballast/blob/main/docs/results-retrieval.md)
|
| 57 |
for generative numbers.*
|
| 58 |
|
| 59 |
-

|
| 60 |
-
|
| 61 |
-
|
|
|
|
|
|
|
|
|
|
| 62 |
|
| 63 |
The pattern replicates across two unrelated families: raw floors spread,
|
| 64 |
ballasted ceilings compress — and the **ballasted Qwen 4B beats the ballasted
|
|
|
|
| 56 |
[retrieval results](https://github.com/OpenBallast/ballast/blob/main/docs/results-retrieval.md)
|
| 57 |
for generative numbers.*
|
| 58 |
|
| 59 |
+

|
| 60 |
+

|
| 61 |
+
*Both families, bf16, oracle-composed (perfect entity linking). The measured
|
| 62 |
+
real-linker band realizes ~63% of the corpus curves' gain. In the Qwen chart
|
| 63 |
+
the 0.8B + corpus curve ends above every bare model in the family — including
|
| 64 |
+
the 9B at 6× its footprint.*
|
| 65 |
|
| 66 |
The pattern replicates across two unrelated families: raw floors spread,
|
| 67 |
ballasted ceilings compress — and the **ballasted Qwen 4B beats the ballasted
|