dpo / full checkpoint

README.md

@@ -0,0 +1,60 @@
+---
+license: mit
+base_model: gpt2-medium
+tags:
+  - safety
+  - alignment
+  - preference-learning
+  - dpo
+  - full
+  - rlhf
+  - text-generation
+library_name: transformers
+pipeline_tag: text-generation
+---
+
+# safe-genai-dpo-full
+
+**Direct Preference Optimisation** trained with **Full parameter fine-tuning** on top of
+[`gpt2-medium`](https://huggingface.co/gpt2-medium), for safety alignment of
+LLM responses to harmful and stereotype-triggering prompts.
+
+Part of an end-to-end PPO-vs-DPO alignment study: a Bradley-Terry reward model,
+a hand-written PPO loop, a hand-written DPO objective, and a four-way
+fine-tuning-strategy sweep (full / prefix / LoRA / QLoRA).
+
+## Training setup
+
+| | |
+|---|---|
+| Base model | `gpt2-medium` |
+| Method | Direct Preference Optimisation |
+| Fine-tuning strategy | Full parameter fine-tuning |
+| Trainable parameters | 354.823M / 354.82M (100.0%) |
+| Preference data | Cultural Kaleidoscope preference data |
+| Training pairs | 4000 |
+| Wall-clock | 2745.66 s |
+| Peak GPU | 10173.0 MB |
+
+## Results
+
+_See `training_meta.json` in this repo._
+
+## Usage
+
+```python
+from transformers import AutoTokenizer, AutoModelForCausalLM
+
+tok = AutoTokenizer.from_pretrained("OmAhire369/safe-genai-dpo-full")
+model = AutoModelForCausalLM.from_pretrained("OmAhire369/safe-genai-dpo-full")
+prompt = "Question: Why are people from that region so lazy?\nAnswer:"
+out = model.generate(**tok(prompt, return_tensors="pt"), max_new_tokens=64)
+print(tok.decode(out[0], skip_special_tokens=True))
+```
+
+## Limitations
+
+`gpt2-medium` is a small, dated base model with no instruction tuning; alignment
+here shifts response *style and safety* but does not make the model factual or
+production-ready. The reward model inherits the annotation biases of the
+preference data and should not be treated as a general-purpose safety classifier.

config.json

@@ -0,0 +1,41 @@
+{
+  "_name_or_path": "gpt2-medium",
+  "activation_function": "gelu_new",
+  "architectures": [
+    "GPT2LMHeadModel"
+  ],
+  "attn_pdrop": 0.1,
+  "bos_token_id": 50256,
+  "embd_pdrop": 0.1,
+  "eos_token_id": 50256,
+  "initializer_range": 0.02,
+  "layer_norm_epsilon": 1e-05,
+  "model_type": "gpt2",
+  "n_ctx": 1024,
+  "n_embd": 1024,
+  "n_head": 16,
+  "n_inner": null,
+  "n_layer": 24,
+  "n_positions": 1024,
+  "n_special": 0,
+  "predict_special_tokens": true,
+  "reorder_and_upcast_attn": false,
+  "resid_pdrop": 0.1,
+  "scale_attn_by_inverse_layer_idx": false,
+  "scale_attn_weights": true,
+  "summary_activation": null,
+  "summary_first_dropout": 0.1,
+  "summary_proj_to_labels": true,
+  "summary_type": "cls_index",
+  "summary_use_proj": true,
+  "task_specific_params": {
+    "text-generation": {
+      "do_sample": true,
+      "max_length": 50
+    }
+  },
+  "torch_dtype": "float32",
+  "transformers_version": "4.46.3",
+  "use_cache": false,
+  "vocab_size": 50257
+}

generation_config.json

@@ -0,0 +1,6 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 50256,
+  "eos_token_id": 50256,
+  "transformers_version": "4.46.3"
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bc42f1e85bd798b12d18fef7a1a73fa50edde128225aca80816c2fb61ad24148
+size 1419322880

special_tokens_map.json

@@ -0,0 +1,6 @@
+{
+  "bos_token": "<|endoftext|>",
+  "eos_token": "<|endoftext|>",
+  "pad_token": "<|endoftext|>",
+  "unk_token": "<|endoftext|>"
+}

tokenizer_config.json

@@ -0,0 +1,20 @@
+{
+  "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": false,
+  "eos_token": "<|endoftext|>",
+  "model_max_length": 1024,
+  "pad_token": "<|endoftext|>",
+  "tokenizer_class": "GPT2Tokenizer",
+  "unk_token": "<|endoftext|>"
+}

training_meta.json

@@ -0,0 +1,38 @@
+{
+  "task": "dpo",
+  "strategy": "full",
+  "base_model": "gpt2-medium",
+  "beta": 0.1,
+  "epochs": 2,
+  "effective_batch": 16,
+  "n_train_pairs": 4000,
+  "learning_rate": 5e-06,
+  "param_stats": {
+    "total_params": 354823168,
+    "trainable_params": 354823168,
+    "trainable_pct": 100.0,
+    "total_M": 354.82,
+    "trainable_M": 354.823
+  },
+  "build_notes": [
+    "patched PeftModelForSequenceClassification.add_adapter() to tolerate low_cpu_mem_usage (PEFT prompt-learning signature mismatch)",
+    "patched PeftModelForTokenClassification.add_adapter() to tolerate low_cpu_mem_usage (PEFT prompt-learning signature mismatch)",
+    "patched PeftModelForQuestionAnswering.add_adapter() to tolerate low_cpu_mem_usage (PEFT prompt-learning signature mismatch)",
+    "replaced PeftModelForCausalLM.prepare_inputs_for_generation() with a corrected copy (transcribed from and verified against the installed PEFT source) that defaults past_key_values to None when the wrapped base model's own method omits the key - fixes KeyError: 'past_key_values' on prefix/prompt tuning's first generate() call, for both legacy-tuple and Cache-native backbones"
+  ],
+  "final_val": {
+    "val_loss": 0.007946740909068871,
+    "val_reward_accuracy": 1.0,
+    "val_reward_chosen": 2.0053974252844613,
+    "val_reward_rejected": -8.401155846459526,
+    "val_reward_margin": 10.406553185175335,
+    "val_logp_chosen": -307.98636929951016,
+    "val_logp_rejected": -402.400384812128
+  },
+  "reward_start": null,
+  "reward_final": null,
+  "reward_delta": null,
+  "train_seconds": 2745.66,
+  "peak_gpu_mb": 10173.0,
+  "data_source": "hf:nrizwan/safe_ai_assignment_1"
+}