--- license: mit base_model: - meta-llama/Llama-3.2-1B-Instruct library_name: transformers --- # Llama-3.2-0.5B-Instruct-small This is a small version of [meta-llama/Llama-3.2-1B-Instruct](https://huggingface.co/meta-llama/Llama-3.2-1B-Instruct) created for testing and development. It preserves the full architecture characteristics of the base model (GQA attention, SwiGLU MLPs, llama3 rope scaling, tied embeddings, identical hyperparameters) with fewer layers. ## Model Details - **Base Model**: meta-llama/Llama-3.2-1B-Instruct - **Architecture**: LlamaForCausalLM (model_type `llama`) - **Total Parameters**: 0.51B (505,956,352) - **Activated Parameters**: 0.51B (dense model) ## Configuration Changes The following parameters were reduced from the original model: | Parameter | Base | Small | |---|---|---| | `num_hidden_layers` | 16 | 4 | All other hyperparameters are identical to the base: hidden_size 2048, intermediate_size 8192, num_attention_heads 32, num_key_value_heads 8, head_dim 64, vocab_size 128256, `tie_word_embeddings` true, rope/llama3 scaling parameters, max_position_embeddings 131072, etc. This is a dense model with a single layer type, so no layer-type or expert changes apply. ## Checkpoint Structure Single `model.safetensors` file with 38 bf16 tensors (the base is also a single file with 146 tensors). Tensor naming and per-layer tensor shapes match the base checkpoint exactly modulo layer count (verified against the base checkpoint's safetensors header; layer-0/embedding/norm shapes identical). As in the base, `lm_head.weight` is absent from the checkpoint because embeddings are tied. ## Usage ```python from transformers import AutoModelForCausalLM, AutoTokenizer model = AutoModelForCausalLM.from_pretrained("inference-optimization/Llama-3.2-0.5B-Instruct-small", device_map="auto") tokenizer = AutoTokenizer.from_pretrained("inference-optimization/Llama-3.2-0.5B-Instruct-small") input_ids = tokenizer("According to all known laws", return_tensors="pt").input_ids.to(model.device) output = model.generate(input_ids, max_new_tokens=20) print(tokenizer.decode(output[0])) ``` ## Creation Process This model was created using the llm-compressor `create-tiny-model` skill. Built from the base `AutoConfig` with only the allowed change above (layer count), initialized with `init_weights()` (bf16), and fine-tuned on a toy copypasta dataset until perplexity ~1.0 (target 3.0). ## Validation ``` $ python validate_tiny_model.py Llama-3.2-0.5B-Instruct-small-ft Success: 1.0484492778778076 <= 10.0 Generating sample text: According to all known laws of aviation, there is no way a bee should be able to fly ``` ## Notes - Dense model: no MoE experts, no per-layer type lists; the only allowed change needed to shrink it is the layer count. - `tie_word_embeddings` kept true (as in base); the checkpoint therefore contains no separate `lm_head` tensor, matching the base layout. - The base model repository is gated; this small version is published under the `inference-optimization` org for testing convenience.