Add RL v2 step20 manifest
Browse files
rl_v2_helpsteer_step20/manifest.json
ADDED
|
@@ -0,0 +1,34 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"folder": "rl_v2_helpsteer_step20",
|
| 3 |
+
"repo": "Asilarkness/testgeniy",
|
| 4 |
+
"checkpoint": "/marimo/testgeniy_rl_real/checkpoints/real_rl_v2_step_0020.pt",
|
| 5 |
+
"version": "testgeniy-v6-real-online-grpo-helpsteer-v2-step20",
|
| 6 |
+
"base_checkpoint": "/marimo/testgeniy_v6_clean/checkpoints/dialogue_sft.pt",
|
| 7 |
+
"real_only_prompts": true,
|
| 8 |
+
"synthetic_prompts": false,
|
| 9 |
+
"mtp": false,
|
| 10 |
+
"algorithm": "online_group_relative_policy_gradient",
|
| 11 |
+
"human_reward": {
|
| 12 |
+
"dataset": "nvidia/HelpSteer2",
|
| 13 |
+
"validation": "OpenAssistant/oasst1 preferences"
|
| 14 |
+
},
|
| 15 |
+
"fixed_suite": {
|
| 16 |
+
"gsm8k_test": 24.0,
|
| 17 |
+
"math500_test": 8.0,
|
| 18 |
+
"arc_challenge_test": 26.0,
|
| 19 |
+
"folio_validation": 29.0,
|
| 20 |
+
"composite": 21.75
|
| 21 |
+
},
|
| 22 |
+
"baseline_v6": {
|
| 23 |
+
"gsm8k_test": 24.0,
|
| 24 |
+
"math500_test": 7.0,
|
| 25 |
+
"arc_challenge_test": 26.0,
|
| 26 |
+
"folio_validation": 29.0,
|
| 27 |
+
"composite": 21.5
|
| 28 |
+
},
|
| 29 |
+
"human_holdouts": {
|
| 30 |
+
"oasst_validation_accuracy": 55.0,
|
| 31 |
+
"helpsteer_wide500_accuracy": 41.6
|
| 32 |
+
},
|
| 33 |
+
"status": "best_measured_candidate_not_primary"
|
| 34 |
+
}
|