Tijmen2 commited on
Commit
800a5ef
1 Parent(s): b49df56

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +15 -2
README.md CHANGED
@@ -39,9 +39,10 @@ device = torch.device("cuda")
39
  model = AutoModelForCausalLM.from_pretrained(model_path).to(device, dtype=torch.bfloat16)
40
  tokenizer = AutoTokenizer.from_pretrained(model_path)
41
 
42
- def ask_cosmosage(question, answer_start=''):
43
- prompt = f"You are cosmosage, an AI programmed to be a cosmology expert. You answer the USER's question clearly in long form, always providing context. When appropriate, provide a reference.USER: {question}ASSISTANT: {answer_start}"
44
  input_ids = tokenizer.encode(prompt, return_tensors="pt").to(device)
 
45
  generated_ids = model.generate(input_ids, max_length=1024, do_sample=True, temperature=0.7, top_k=None, pad_token_id=tokenizer.eos_token_id)
46
  generated_text = tokenizer.decode(generated_ids[0], skip_special_tokens=True)
47
  answer = generated_text.split("ASSISTANT:")[-1]
@@ -125,6 +126,18 @@ heavy specialization in cosmology has not come at much of a cost on general-purp
125
  is GSM8k, which is a collection of grade school math problems. Here, cosmosage performs significantly better
126
  than OpenHermes-2.5-Mistral-7B.
127
 
 
 
 
 
 
 
 
 
 
 
 
 
128
  ## Example output
129
 
130
  **User:**
 
39
  model = AutoModelForCausalLM.from_pretrained(model_path).to(device, dtype=torch.bfloat16)
40
  tokenizer = AutoTokenizer.from_pretrained(model_path)
41
 
42
+ def ask_cosmosage(question):
43
+ prompt = f"You are cosmosage, an AI programmed to provide excellent and detailed answers to the user's question. You are an expert cosmology assistant, able to answer questions on the cosmic microwave background, galaxy formation, large scale structure, theoretical cosmology, inflation, big bang nucleosynthesis, cosmology instrumentation, and other related topics. Please assume the user is fluent in scientific terminology. Elaborate where possible to give a complete answer. If you do not know, say you do not know.▁ USER: {question}ASSISTANT:"
44
  input_ids = tokenizer.encode(prompt, return_tensors="pt").to(device)
45
+ print(input_ids)
46
  generated_ids = model.generate(input_ids, max_length=1024, do_sample=True, temperature=0.7, top_k=None, pad_token_id=tokenizer.eos_token_id)
47
  generated_text = tokenizer.decode(generated_ids[0], skip_special_tokens=True)
48
  answer = generated_text.split("ASSISTANT:")[-1]
 
126
  is GSM8k, which is a collection of grade school math problems. Here, cosmosage performs significantly better
127
  than OpenHermes-2.5-Mistral-7B.
128
 
129
+ ## Instruction format
130
+
131
+ cosmosage_v2 was trained with the "inst" chat template as implemented in axolotl v0.4.0. This resulted in an
132
+ unusual instruction format:
133
+ ```raw
134
+ <s> {system prompt}▁ USER: {question}▁ ASSISTANT:
135
+ ```
136
+ Note the use of the U+2581 Lower One Eighth Block Unicode Character to separate the different sections. The
137
+ example code in the Usage section above correctly implements this format.
138
+
139
+ Fortunately, cosmosage_v2 does not appear to be too sensitive to deviations from this format.
140
+
141
  ## Example output
142
 
143
  **User:**