Push model using huggingface_hub.

README.md

@@ -26,7 +26,7 @@ You can then generate text as follows:
 ```python
 from transformers import pipeline
 
-generator = pipeline("text-generation", model="jvelja//tmp/tmp1th6hrkz/jvelja/gemma2b-sanity-multivllm_0")
+generator = pipeline("text-generation", model="jvelja//tmp/tmptn97qlzw/jvelja/gemma2b-sanity-multivllm_0")
 outputs = generator("Hello, my llama is cute")
 ```
 
@@ -36,8 +36,8 @@ If you want to use the model for training or to obtain the outputs from the valu
 from transformers import AutoTokenizer
 from trl import AutoModelForCausalLMWithValueHead
 
-tokenizer = AutoTokenizer.from_pretrained("jvelja//tmp/tmp1th6hrkz/jvelja/gemma2b-sanity-multivllm_0")
-model = AutoModelForCausalLMWithValueHead.from_pretrained("jvelja//tmp/tmp1th6hrkz/jvelja/gemma2b-sanity-multivllm_0")
+tokenizer = AutoTokenizer.from_pretrained("jvelja//tmp/tmptn97qlzw/jvelja/gemma2b-sanity-multivllm_0")
+model = AutoModelForCausalLMWithValueHead.from_pretrained("jvelja//tmp/tmptn97qlzw/jvelja/gemma2b-sanity-multivllm_0")
 
 inputs = tokenizer("Hello, my llama is cute", return_tensors="pt")
 outputs = model(**inputs, labels=inputs["input_ids"])

adapter_model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:afaca8044f042ff38b35e1ddf38af6001eea7bd54e3154c207b7f2bf2ca76204
+oid sha256:22a1c2bc4f56acaed4c390887952f5317c32c53d412f47a88b73cdac053c7320
 size 12793376

config.json

@@ -1,8 +1,8 @@
 {
   "accelerator_kwargs": {},
   "adap_kl_ctrl": true,
-  "backward_batch_size": 64,
-  "batch_size": 64,
+  "backward_batch_size": 128,
+  "batch_size": 128,
   "cliprange": 0.2,
   "cliprange_value": 0.2,
   "compare_steps": 1,
@@ -11,12 +11,12 @@
   "exp_name": "stego_trainer",
   "forward_batch_size": null,
   "gamma": 1,
-  "global_backward_batch_size": 64,
-  "global_batch_size": 64,
+  "global_backward_batch_size": 128,
+  "global_batch_size": 128,
   "gradient_accumulation_steps": 4,
   "gradient_checkpointing": false,
   "horizon": 10000,
-  "init_kl_coef": 0.2,
+  "init_kl_coef": 0.05,
   "is_encoder_decoder": false,
   "is_peft_model": true,
   "kl_penalty": "kl",
@@ -24,7 +24,7 @@
   "learning_rate": 2e-05,
   "log_with": "wandb",
   "max_grad_norm": null,
-  "mini_batch_size": 16,
+  "mini_batch_size": 32,
   "model_name": "unsloth/gemma-2-2b-it",
   "optimize_cuda_cache": true,
   "optimize_device_cache": false,
@@ -38,13 +38,13 @@
   "score_clip": null,
   "seed": 0,
   "steps": 20000,
-  "target": 6,
+  "target": 12.0,
   "target_kl": 1,
   "task_name": null,
   "tracker_kwargs": {
     "wandb": {
-      "name": "cv_gemma-2-2b-it_to_distilbert-base-uncased_EBS64_Joan",
-      "notes": "Dataset: cv\n    Same Prompt:  \n    Payload Prefixes: ['Movie Review: This movie was really amazing!', 'Movie Review: This movie was really terrible!']\n    Payload Template: Movie Review: This movie was really {payload}!\n    Separate Enc/Dec Data: True\n\n    Encoder: gemma-2-2b-it (LR: 2e-05)\n    Decoder: distilbert-base-uncased (LR: 2e-05)\n    Train Loop: v2_dylan\n\n    Effective Batch Sizes:\n    - Encoder: 64\n    - Decoder: 256\n\n    Training Iterations:\n    - Encoder updates: 80\n    - Decoder updates: 240\n    - Update Encoder First: False\n\n    Temperatures:\n    - Decoder Training: 1.0\n    - Encoder Training: 1.0\n    - Evaluation: 1.0\n\n    Encoder Parameters:\n    - KL Coefficient: 0.1\n    - LoRA: True\n    - Quantization: False\n    - Output Length: {'min': 50, 'max': 56}\n\n    Decoder Parameters:\n    - New Classification Head: True\n    - Use Probs Reward: False\n    - Weight Decay: 0.01\n    - Update Parameters: {'head': True, 'body': True}\n\n    Training Configuration:\n    - Update Encoder: True\n    - Update Decoder: True\n    - Paraphrase: False\n    - Leak Password: False\n    - WandB Logging: True\n    - Eval Every N: 50\n    - Number of Epochs: 100000\n\n    Debug:\n    - Override Dec Batch: False",
+      "name": "cv_gemma-2-2b-it_to_distilbert-base-uncased_EBS128_Joan",
+      "notes": "Dataset: cv\n    Same Prompt:  \n    Payload Prefixes: ['Movie Review: This movie was really amazing!', 'Movie Review: This movie was really terrible!']\n    Payload Template: Movie Review: This movie was really {payload}!\n    Separate Enc/Dec Data: True\n\n    Encoder: gemma-2-2b-it (LR: 2e-05)\n    Decoder: distilbert-base-uncased (LR: 2e-05)\n    Train Loop: v2_dylan\n\n    Effective Batch Sizes:\n    - Encoder: 128\n    - Decoder: 256\n\n    Training Iterations:\n    - Encoder updates: 80\n    - Decoder updates: 240\n    - Update Encoder First: False\n\n    Temperatures:\n    - Decoder Training: 1.0\n    - Encoder Training: 1.0\n    - Evaluation: 1.0\n\n    Encoder Parameters:\n    - KL Coefficient: 0.05\n    - LoRA: True\n    - Quantization: False\n    - Output Length: {'min': 50, 'max': 56}\n\n    Decoder Parameters:\n    - New Classification Head: True\n    - Use Probs Reward: False\n    - Weight Decay: 0.01\n    - Update Parameters: {'head': True, 'body': True}\n\n    Training Configuration:\n    - Update Encoder: True\n    - Update Decoder: True\n    - Paraphrase: False\n    - Leak Password: False\n    - WandB Logging: True\n    - Eval Every N: 50\n    - Number of Epochs: 100000\n\n    Debug:\n    - Override Dec Batch: False",
       "tags": [
         "cv",
         "gemma-2-2b-it",
@@ -52,7 +52,7 @@
         "v2_dylan",
         "enc_lr_2e-05",
         "dec_lr_2e-05",
-        "enc_eff_bs_64",
+        "enc_eff_bs_128",
         "dec_eff_bs_256",
         "enc_updates_80",
         "dec_updates_240",

pytorch_model.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:78e2211ed615bb005af73ea05a94170310ad35e2f7a7f98ef6f832fbf3d4677a
+oid sha256:af3a50393088d7b124587b883915ea2a75a4e8e0d8f735541c6c333cfa80b498
 size 10748