Upload README.md with huggingface_hub
Browse files
README.md
CHANGED
|
@@ -30,6 +30,9 @@ trajectory can be re-scored offline without a GPU or a model download.
|
|
| 30 |
| `sft3_hard30/`, `sft3_hard90/`, `sft3_eval/` | round-2 adapter (mix2) |
|
| 31 |
| `sft4_hard30/`, `sft4_hard90/`, `sft4_eval/` | **round-3 adapter, the one that ships** |
|
| 32 |
| `sft4_hard90_hardened/` | shipped adapter on clean hard-90 with the hardened system prompt (86/90 vs 84/90 default) |
|
|
|
|
|
|
|
|
|
|
| 33 |
| `sft5_hard30/`, `sft5_hard90/`, `sft5_eval/`, `sft5_sanity/` | mixture ablation (fewer ask rows) |
|
| 34 |
| `sft3_seed7_hard30/`, `sft3_seed7_hard90/`, `sft3_seed123_hard30/`, `sft3_seed123_hard90/` | seed control on the mix2 recipe |
|
| 35 |
| `sft4_seed7_hard30/`, `sft4_seed7_hard90/`, `sft4_seed123_hard30/`, `sft4_seed123_hard90/` | mix3 seed control (seeds 7 and 123) for the shipped seed 42 |
|
|
|
|
| 30 |
| `sft3_hard30/`, `sft3_hard90/`, `sft3_eval/` | round-2 adapter (mix2) |
|
| 31 |
| `sft4_hard30/`, `sft4_hard90/`, `sft4_eval/` | **round-3 adapter, the one that ships** |
|
| 32 |
| `sft4_hard90_hardened/` | shipped adapter on clean hard-90 with the hardened system prompt (86/90 vs 84/90 default) |
|
| 33 |
+
| `sft4_shift_hard30/` | shipped adapter on hard-30 with every tool and parameter renamed (25/30 vs 29/30 canonical) |
|
| 34 |
+
| `sft4_devanagari30/` | shipped adapter on the Devanagari version of hard-30, canonical schema (25/30) |
|
| 35 |
+
| `base_shift_hard30/` | untuned base under the same rename (0/30, loop collapse 0.60) |
|
| 36 |
| `sft5_hard30/`, `sft5_hard90/`, `sft5_eval/`, `sft5_sanity/` | mixture ablation (fewer ask rows) |
|
| 37 |
| `sft3_seed7_hard30/`, `sft3_seed7_hard90/`, `sft3_seed123_hard30/`, `sft3_seed123_hard90/` | seed control on the mix2 recipe |
|
| 38 |
| `sft4_seed7_hard30/`, `sft4_seed7_hard90/`, `sft4_seed123_hard30/`, `sft4_seed123_hard90/` | mix3 seed control (seeds 7 and 123) for the shipped seed 42 |
|