Upload RLHF checkpoint at epoch 6

README.md

@@ -0,0 +1,45 @@
+# f9bs90mv-rlhf-checkpoint-gpt-neo-125m-irl-epoch-6
+
+This is a RLHF model checkpoint trained at epoch 6.
+
+## Model Information
+
+- **Base Model**: EleutherAI/gpt-neo-125M
+- **Reward Type**: irl
+- **Dataset**: allenai/real-toxicity-prompts
+- **Training Epoch**: 6
+
+
+## IRL Configuration
+
+- **Likelihood Type**: bradley_terry
+- **Normalization Strategy**: none
+- **IRL Artifact**: matthieubou-imperial-college-london/bayes_irl_vi/posterior_bradley_terry_rkiq5pd8:v0
+- **Use Raw Score**: True
+
+## Usage
+
+This checkpoint can be loaded using the HuggingFace Transformers library:
+
+```python
+from transformers import AutoModelForCausalLM
+from trl import AutoModelForCausalLMWithValueHead
+
+# Load the checkpoint
+model = AutoModelForCausalLMWithValueHead.from_pretrained("MattBou00/f9bs90mv-rlhf-checkpoint-gpt-neo-125m-irl-epoch-6")
+```
+
+## Training Configuration
+
+The training configuration is saved in `training_config.yaml`.
+
+---
+language: en
+tags:
+- rlhf
+- checkpoint
+- irl
+- gpt-neo-125m
+library_name: transformers
+pipeline_tag: text-generation
+---

config.json

@@ -0,0 +1,54 @@
+{
+  "activation_function": "gelu_new",
+  "architectures": [
+    "GPTNeoForCausalLM"
+  ],
+  "attention_dropout": 0,
+  "attention_layers": [
+    "global",
+    "local",
+    "global",
+    "local",
+    "global",
+    "local",
+    "global",
+    "local",
+    "global",
+    "local",
+    "global",
+    "local"
+  ],
+  "attention_types": [
+    [
+      [
+        "global",
+        "local"
+      ],
+      6
+    ]
+  ],
+  "bos_token_id": 50256,
+  "classifier_dropout": 0.1,
+  "embed_dropout": 0,
+  "eos_token_id": 50256,
+  "gradient_checkpointing": false,
+  "hidden_size": 768,
+  "initializer_range": 0.02,
+  "intermediate_size": null,
+  "layer_norm_epsilon": 1e-05,
+  "max_position_embeddings": 2048,
+  "model_type": "gpt_neo",
+  "num_heads": 12,
+  "num_layers": 12,
+  "resid_dropout": 0,
+  "summary_activation": null,
+  "summary_first_dropout": 0.1,
+  "summary_proj_to_labels": true,
+  "summary_type": "cls_index",
+  "summary_use_proj": true,
+  "torch_dtype": "float32",
+  "transformers_version": "4.54.0",
+  "use_cache": true,
+  "vocab_size": 50257,
+  "window_size": 256
+}

generation_config.json

@@ -0,0 +1,6 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 50256,
+  "eos_token_id": 50256,
+  "transformers_version": "4.54.0"
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0c2c1eb3959bb8c134a90ffc7cff96a0ddb73e7182d7132cb07108c565b187a3
+size 500814588

special_tokens_map.json

@@ -0,0 +1,24 @@
+{
+  "bos_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": "<|endoftext|>",
+  "unk_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  }
+}

tokenizer_config.json

@@ -0,0 +1,23 @@
+{
+  "add_bos_token": false,
+  "add_prefix_space": false,
+  "added_tokens_decoder": {
+    "50256": {
+      "content": "<|endoftext|>",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "bos_token": "<|endoftext|>",
+  "clean_up_tokenization_spaces": true,
+  "eos_token": "<|endoftext|>",
+  "errors": "replace",
+  "extra_special_tokens": {},
+  "model_max_length": 2048,
+  "pad_token": "<|endoftext|>",
+  "tokenizer_class": "GPT2Tokenizer",
+  "unk_token": "<|endoftext|>"
+}

training_config.yaml

@@ -0,0 +1,65 @@
+model:
+  name: EleutherAI/gpt-neo-125M
+  reward_type: irl
+  reward_model: matthieubou-imperial-college-london/bayes_irl_vi/posterior_bradley_terry_rkiq5pd8:v0
+  base_model_name: EleutherAI/pythia-410m
+  likelihood_type: bradley_terry
+  use_raw_irl_score: true
+  irl_normalization_strategy: none
+  irl_reward_scale: 1.0
+  n_posterior_samples: 100
+  uncertainty_threshold: 0.45
+  use_distance_sampling: false
+  learning_rate: 2.0e-06
+  batch_size: 128
+  mini_batch_size: 8
+  gradient_accumulation_steps: 16
+  max_sequence_length: 512
+  generation:
+    min_length: -1
+    top_k: 0
+    top_p: 0.8
+    temperature: 0.7
+    do_sample: true
+    output_min_length: 20
+    output_max_length: 30
+rlhf:
+  model:
+    ppo_epochs: 3
+    init_kl_coef: 0.2
+    target: 1.0
+    cliprange: 0.1
+    cliprange_value: 0.2
+    vf_coef: 0.1
+    adap_kl_ctrl: true
+    use_score_norm: true
+    ratio_threshold: 10.0
+training:
+  num_train_epochs: 80
+  save_freq: 2
+  eval_freq: 2
+  seed: 0
+dataset:
+  name: allenai/real-toxicity-prompts
+  toxicity_metric: profanity
+  toxicity_threshold: 0.7
+  input_min_text_length: 10
+  input_max_text_length: 100
+  test_size: 0.1
+  num_samples: 20000
+output:
+  push_to_hub: true
+  push_checkpoints_to_hub: true
+  push_final_model_to_hub: true
+  hub_org: MattBou00
+  repo_name_prefix: rlhf-checkpoint
+  private: false
+wandb:
+  project: irl-rlhf-detox
+  entity: matthieubou-imperial-college-london
+  name: null
+now: 2025-08-03_00-18-13
+logging:
+  use_wandb: true
+  project_name: irl-rlhf-detox
+  wandb_mode: online