verl GRPO trained model at step 1
Browse files- README.md +34 -0
- adapter_config.json +45 -0
- adapter_model.safetensors +3 -0
README.md
ADDED
|
@@ -0,0 +1,34 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
language: en
|
| 3 |
+
license: apache-2.0
|
| 4 |
+
tags:
|
| 5 |
+
- verl
|
| 6 |
+
- grpo
|
| 7 |
+
- math
|
| 8 |
+
- reasoning
|
| 9 |
+
- rl
|
| 10 |
+
- lora
|
| 11 |
+
- peft
|
| 12 |
+
base_model: google/gemma-2-9b-it
|
| 13 |
+
library_name: peft
|
| 14 |
+
---
|
| 15 |
+
|
| 16 |
+
# thejaminator/grpo-feature-vector-step-1
|
| 17 |
+
|
| 18 |
+
This is a LoRA adapter trained using [verl](https://github.com/volcengine/verl) with GRPO (Group Relative Policy Optimization)
|
| 19 |
+
on math reasoning tasks.
|
| 20 |
+
|
| 21 |
+
## Training Details
|
| 22 |
+
|
| 23 |
+
- **Base model**: google/gemma-2-9b-it
|
| 24 |
+
- **Framework**: verl GRPO
|
| 25 |
+
- **Training steps**: 1
|
| 26 |
+
- **Dataset**: Math reasoning problems
|
| 27 |
+
- **Batch size**: 8
|
| 28 |
+
- **Learning rate**: 5e-05
|
| 29 |
+
- **LoRA rank**: 64
|
| 30 |
+
- **LoRA alpha**: 128.0
|
| 31 |
+
- **Number of generations**: 16
|
| 32 |
+
|
| 33 |
+
|
| 34 |
+
Generated from verl LoRA checkpoint: `/workspace/verl_outputs_feature_vector/global_step_1/actor/lora_adapter`
|
adapter_config.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"task_type": "CAUSAL_LM",
|
| 3 |
+
"peft_type": "LORA",
|
| 4 |
+
"auto_mapping": null,
|
| 5 |
+
"base_model_name_or_path": "google/gemma-2-9b-it",
|
| 6 |
+
"revision": null,
|
| 7 |
+
"inference_mode": false,
|
| 8 |
+
"r": 64,
|
| 9 |
+
"target_modules": [
|
| 10 |
+
"gate_proj",
|
| 11 |
+
"k_proj",
|
| 12 |
+
"v_proj",
|
| 13 |
+
"q_proj",
|
| 14 |
+
"down_proj",
|
| 15 |
+
"up_proj",
|
| 16 |
+
"o_proj"
|
| 17 |
+
],
|
| 18 |
+
"exclude_modules": null,
|
| 19 |
+
"lora_alpha": 128.0,
|
| 20 |
+
"lora_dropout": 0.0,
|
| 21 |
+
"fan_in_fan_out": false,
|
| 22 |
+
"bias": "none",
|
| 23 |
+
"use_rslora": false,
|
| 24 |
+
"modules_to_save": null,
|
| 25 |
+
"init_lora_weights": true,
|
| 26 |
+
"layers_to_transform": null,
|
| 27 |
+
"layers_pattern": null,
|
| 28 |
+
"rank_pattern": {},
|
| 29 |
+
"alpha_pattern": {},
|
| 30 |
+
"megatron_config": null,
|
| 31 |
+
"megatron_core": "megatron.core",
|
| 32 |
+
"trainable_token_indices": null,
|
| 33 |
+
"loftq_config": {},
|
| 34 |
+
"eva_config": null,
|
| 35 |
+
"corda_config": null,
|
| 36 |
+
"use_dora": false,
|
| 37 |
+
"use_qalora": false,
|
| 38 |
+
"qalora_group_size": 16,
|
| 39 |
+
"layer_replication": null,
|
| 40 |
+
"runtime_config": {
|
| 41 |
+
"ephemeral_gpu_offload": false
|
| 42 |
+
},
|
| 43 |
+
"lora_bias": false,
|
| 44 |
+
"target_parameters": null
|
| 45 |
+
}
|
adapter_model.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:d709cbb6bc6f70a27f3a8fb788ac6126d661db8baf358be14a769086ae35a0c4
|
| 3 |
+
size 864368248
|