Create README.md

README.md

@@ -0,0 +1,31 @@
+---
+license: apache-2.0
+datasets:
+- allenai/dolma
+---
+# OLMo-Bitnet-1B
+
+OLMo-Bitnet-1B is a 1B parameter model trained using the method described in [The Era of 1-bit LLMs: All Large Language Models are in 1.58 Bits](https://arxiv.org/abs/2402.17764).
+The result of this is that all of the parameter weights take only the values -1, 0, or 1.
+
+It was trained on a 60B subset of the [Dolma](https://huggingface.co/datasets/allenai/dolma) dataset, so it is merely a research proof-of-concept to test out the methodolgy.
+
+A separate training run was run with the exact same hyperparameters, but using standard fp16 weights.
+The comparison can be found in [this wandb report](https://api.wandb.ai/links/emozilla/evltqiv7).
+
+Sample inference code
+```python
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline, TextStreamer
+
+tokenizer = AutoTokenizer.from_pretrained("NousResearch/OLMo-Bitnet-1B")
+model = AutoModelForCausalLM.from_pretrained("NousResearch/OLMo-Bitnet-1B",
+    torch_dtype=torch.bfloat16, trust_remote_code=True, device_map="auto")
+
+streamer = TextStreamer(tokenizer)
+pipe = pipeline("text-generation", model=model, tokenizer=tokenizer, pad_token_id=tokenizer.eos_token_id,
+    temperature=0.8, repetition_penalty=1.1, do_sample=True,streamer=streamer)
+pipe("The capitol of Paris is",  max_new_tokens=256)
+```
+
+Training was performed using [OLMo](https://github.com/allenai/OLMo).