--- license: mit datasets: - HuggingFaceFW/fineweb-edu language: - en pipeline_tag: text-generation tags: - dillion - small - harley-ml - text-generation - transformers - pytorch --- # **Dillion-1.2M** ## **Summary** ``` Task: Text-Generation Total training time: ~2.5 days Inputs: text Outputs: text Params: ~1.2M Framework: PyTorch, transformers Author: Paul Courneya (Harley-ml) ``` ## **Description** Dillion is a 1.2M parameter language model trained on ~9B tokens of FineWeb-edu. Our goal was to make one of the best sub-1.5M parameter LMs through depth (12 layers) and huge overtraining (about 8900 tokens per parameter). Dillion beats or ties with models much larger than itself such as [SupraMini-v4-2M](https://huggingface.co/SupraLabs/Supra-Mini-v4-2M) and [Tenete-8M](https://huggingface.co/Harley-ml/Tenete-8M). ### Why "Dillion"? I was scrolling through Hugging Face and saw GPT-2, the smallest variant. I looked at its download count and saw 16 million. My brain, for some random reason, hallucinated “Dillion.” So I decided to call my next model, no matter the task or size, Dillion. I decided to dig a bit deeper, and after a quick Google Search, I found that “Dillion” is an alternate spelling of the Irish name Dillon, which translates to “loyal” or “faithful.” But let me tell you, this model ain’t loyal or faithful; actually, it probably doesn’t even know what those words mean. ## Architecture Dillion-1.2M uses the Qwen3.5 architecture. | Parameter | Value | | ------------------------- | ---------------- | | `NUM_HIDDEN_LAYERS` | `12` | | `HIDDEN_SIZE` | `72` | | `NUM_ATTENTION_HEADS` | `3` | | `NUM_KEY_VALUE_HEADS` | `3` | | `VOCAB_SIZE` | `3076` | | `INTERMEDIATE_SIZE` | `288` | | `ROPE_THETA` | `10000.0` | | `MAX_POSITION_EMBEDDINGS` | `384` | | `LAYER_TYPES` | `full_attention` | ## Training ### Hardware We trained Dillion for 0.71 epochs on 14B (only saw ~9B) tokens of FineWeb-edu on an RTX 2060 6GB with a batch size of 72 and a gradient accumulation of 4. ### Training Results | epoch | train_loss | train_ppl | train_bpb | eval_loss | eval_ppl | eval_bpb | | ------- | ---------: | --------: | --------: | --------: | -------: | -------: | | 0.02368 | 4.553 | 94.917 | 1.875 | 4.492 | 89.300 | 1.850 | | 0.04736 | 3.958 | 52.353 | 1.630 | 3.943 | 51.573 | 1.624 | | 0.07104 | 3.763 | 43.077 | 1.550 | 3.758 | 42.863 | 1.548 | | 0.09472 | 3.672 | 39.330 | 1.512 | 3.670 | 39.252 | 1.511 | | 0.11840 | 3.620 | 37.338 | 1.491 | 3.620 | 37.338 | 1.491 | | 0.14210 | 3.584 | 36.017 | 1.476 | 3.586 | 36.089 | 1.477 | | 0.16580 | 3.557 | 35.058 | 1.465 | 3.558 | 35.093 | 1.465 | | 0.18940 | 3.538 | 34.398 | 1.457 | 3.536 | 34.329 | 1.456 | | 0.21310 | 3.520 | 33.784 | 1.450 | 3.520 | 33.784 | 1.450 | | 0.23680 | 3.504 | 33.248 | 1.443 | 3.507 | 33.348 | 1.444 | | 0.26050 | 3.494 | 32.917 | 1.439 | 3.494 | 32.917 | 1.439 | | 0.28420 | 3.483 | 32.557 | 1.434 | 3.484 | 32.590 | 1.435 | | 0.30780 | 3.475 | 32.298 | 1.431 | 3.475 | 32.298 | 1.431 | | 0.33150 | 3.465 | 31.976 | 1.427 | 3.468 | 32.073 | 1.428 | | 0.35520 | 3.459 | 31.785 | 1.425 | 3.459 | 31.785 | 1.425 | | 0.37890 | 3.452 | 31.563 | 1.422 | 3.454 | 31.627 | 1.423 | | 0.40260 | 3.445 | 31.343 | 1.419 | 3.447 | 31.406 | 1.420 | | 0.42620 | 3.441 | 31.218 | 1.417 | 3.441 | 31.218 | 1.417 | | 0.44990 | 3.437 | 31.094 | 1.416 | 3.437 | 31.094 | 1.416 | | 0.47360 | 3.431 | 30.908 | 1.413 | 3.433 | 30.969 | 1.414 | | 0.49730 | 3.426 | 30.753 | 1.411 | 3.428 | 30.815 | 1.412 | | 0.52100 | 3.423 | 30.661 | 1.410 | 3.424 | 30.692 | 1.410 | | 0.54460 | 3.419 | 30.539 | 1.408 | 3.420 | 30.569 | 1.409 | | 0.56830 | 3.417 | 30.478 | 1.407 | 3.416 | 30.447 | 1.407 | | 0.59200 | 3.413 | 30.356 | 1.406 | 3.413 | 30.356 | 1.406 | | 0.61570 | 3.409 | 30.235 | 1.404 | 3.410 | 30.265 | 1.404 | | 0.63940 | 3.404 | 30.084 | 1.402 | 3.407 | 30.175 | 1.403 | | 0.66300 | 3.403 | 30.054 | 1.402 | 3.403 | 30.054 | 1.402 | | 0.68670 | 3.397 | 29.874 | 1.399 | 3.401 | 29.994 | 1.401 | ## Benchmarks | Model | Parameters | | --------------- | ---------- | | Dillion | 1,281,384 | | SupraMini-v4-2M | 8,293,888 | | Tenete-8M | 2,623,104 | | Task | Metric | Dillion | SupraMini-v4-2M | Tenete-8M | | -------- | --------------- | -------: | --------------: | ---------: | | ARC Easy | acc_norm | 31.36% | 31.50% | 31.94% | | BLiMP | acc | 62.94% | 60.70% | — | | PiQA | acc_norm | 53.10% | 51.90% | 55.71% | | SWAG | acc_norm | 30.36% | — | 32.97% | | WikiText | bits_per_byte | 1.6161 | — | — | | WikiText | byte_perplexity | 3.0655 | 3.1652 | — | See the raw output from LM Harnes for Dillion [here](https://huggingface.co/Harley-ml/Dillion-1.2M/blob/main/benchmarks.txt) ## Generation Examples Prompt: `The` Output: ``` Twitter and Freees of Brooklyn Press, Oxford University. The Home Council of the Monthly Landing Foundation is a partner with the Great War in the South. The Eighteenth Century has been held on the River Battalion by the Vietnam War, which was laid down by the German Empire to the Nazis. Its first-year period was born on May 1, 1846. ``` --- Prompt: `Artificial Intelligence is` Output: ``` a new technology that has been used to make the processes we use. The Mexican War: Since Ireland, it’s not just one of the most important technologies in America, it can be found in Europe and Japan. The Economics Center for Natural Resources (EU), which was created by Berlin, has become an essential component of its development. Firstly, it will enable the Congress to have the opportunity to create such a system and to generate a great range of resources as well. It also uses a variety of methods to provide more detailed information. Listen to our article on these tools: - Published on 2017-2015 ``` --- Prompt: `I was walking down the street and saw a` Output: ``` balloon on the ground. Before you see that the floor, we started to build a large-scale planetary traffic, which makes it possible for people to move from a magma to a hospital when they were picked up in the shore of the first day. They had a small window on their nests and so much fine space into the roof. Then in the gap between them and the width of the tropical Solar Systems. Many scientists have found that the densely grown snowflakes are being born with the mouth of their own. But there is no evidence of the difference in this condition. The findings are not necessarily an effective way to prevent the spread of the knees and its use as well as other conditions. It's a major issue about how many thousands of molecules will be released ``` ## Use Cases 1. Educational work and research 2. Fine-tuning for downstream use 2. Deployment on edge devices 4. Or for fun. ## Limitations Umm... What do you think? Yeah, everything. But... more speciifcally (yep I splet that wrong; what are you gonna do about it!?) 1. Cannot chat, reason, code, or answer questions 2. Always unfactual 3. No long context handling ## Inference ```python #!/usr/bin/env python3 # ============================================================================= # Inference # ============================================================================= MODEL_DIR = "Harley-ml/Dillion-1.2M" TOKENIZER_PATH = MODEL_DIR # --- Generation settings --- PROMPT = "The" MAX_NEW_TOKENS = 362 TEMPERATURE = 0.6 TOP_P = 0.95 TOP_K = 30 REPETITION_PENALTY = 1.2 DO_SAMPLE = True # ============================================================================= import os import torch from pathlib import Path from transformers import ( AutoModelForCausalLM, AutoTokenizer, PreTrainedTokenizerFast, AddedToken, ) # --------------------------------------------------------------------------- # Device # --------------------------------------------------------------------------- device = ( "cuda" if torch.cuda.is_available() else "mps" if torch.backends.mps.is_available() else "cpu" ) print(f"Device : {device}") # --------------------------------------------------------------------------- # Tokenizer # --------------------------------------------------------------------------- def load_tokenizer(path_or_repo: str): p = Path(path_or_repo) # Case 1: explicit local tokenizer.json file if p.exists() and p.is_file() and p.suffix.lower() == ".json": tok = PreTrainedTokenizerFast(tokenizer_file=str(p.resolve())) # Case 2: local directory or HF repo ID else: tok = AutoTokenizer.from_pretrained(path_or_repo, use_fast=True) # Ensure required special tokens exist if tok.bos_token is None: tok.add_special_tokens({"bos_token": "<|bos|>"}) if tok.eos_token is None: tok.add_special_tokens({"eos_token": "<|eos|>"}) if tok.unk_token is None: tok.add_special_tokens({"unk_token": "<|unk|>"}) if tok.pad_token is None: tok.pad_token = tok.eos_token if tok.eos_token is not None else "<|pad|>" tok.padding_side = "left" return tok print("Loading tokenizer...") tokenizer = load_tokenizer(TOKENIZER_PATH) print(f" Vocab size : {len(tokenizer)}") print(f" BOS : {tokenizer.bos_token!r}") print(f" EOS : {tokenizer.eos_token!r}") print(f" PAD : {tokenizer.pad_token!r} (id={tokenizer.pad_token_id})") # --------------------------------------------------------------------------- # Model # --------------------------------------------------------------------------- print(f"\nLoading model from {MODEL_DIR} ...") model = AutoModelForCausalLM.from_pretrained( MODEL_DIR, torch_dtype=torch.float16 if device == "cuda" else torch.float32, low_cpu_mem_usage=True, ) model.eval() model.to(device) # Safer inference for cache-related issues model.config.use_cache = False if hasattr(model, "generation_config") and model.generation_config is not None: model.generation_config.use_cache = False total_params = sum(p.numel() for p in model.parameters()) print(f" Parameters : {total_params:,}") # --------------------------------------------------------------------------- # Generation helper # --------------------------------------------------------------------------- def generate( prompt: str = PROMPT, max_new_tokens: int = MAX_NEW_TOKENS, temperature: float = TEMPERATURE, top_p: float = TOP_P, top_k: int = TOP_K, repetition_penalty: float = REPETITION_PENALTY, do_sample: bool = DO_SAMPLE, ) -> str: bos = tokenizer.bos_token or "" full_prompt = bos + prompt inputs = tokenizer( full_prompt, return_tensors="pt", add_special_tokens=False, ).to(device) inputs.pop("token_type_ids", None) gen_kwargs = dict( max_new_tokens=max_new_tokens, do_sample=do_sample, repetition_penalty=repetition_penalty, eos_token_id=tokenizer.eos_token_id, pad_token_id=tokenizer.pad_token_id, use_cache=False, ) if do_sample: gen_kwargs["temperature"] = temperature gen_kwargs["top_p"] = top_p gen_kwargs["top_k"] = top_k with torch.inference_mode(): output_ids = model.generate(**inputs, **gen_kwargs) prompt_len = inputs["input_ids"].shape[-1] new_ids = output_ids[0][prompt_len:] return tokenizer.decode(new_ids, skip_special_tokens=True) # --------------------------------------------------------------------------- # Run # --------------------------------------------------------------------------- if __name__ == "__main__": print(f"\nPrompt : {PROMPT!r}") print("-" * 60) output = generate(PROMPT) print("Generated:") print(output) ```