Skip to content

A LLM trained only on data from certain time periods to reduce modern bias

Python

Revise README with language options and model updates

Hayk Grigorian015004a1412a
diff --git a/README.md b/README.md
index 76328b9..88fd6ce 100644
--- a/README.md
+++ b/README.md
@@ -52,7 +52,7 @@ v2 built on llamaforcausallm
Early prompts show the model responding with 1800's language and behavior.
Example: Prompt: "Who art Henry?" and it replied "I know that man, I have did not a black, the storm."
-![TimeLockLLM Sample Output](https://github.com/haykgrigo3/TimeCapsuleLLM/blob/main/london_1800_1850_v0/timelockllm_sample_output.png?raw=true)
+![TimeCapsuleLLM Sample Output](https://github.com/haykgrigo3/TimeCapsuleLLM/blob/main/london_1800_1850_v0/timelockllm_sample_output.png?raw=true)
- No mention of modern concepts
- Mostly era-accurate vocabulary
@@ -65,7 +65,7 @@ A significant improvement over v0.
- Still high factual hallucination rate
- OCR noise (“Digitized by Google”) still present in outputs
-![TimeLockLLM Sample Output](https://github.com/haykgrigo3/TimeCapsuleLLM/blob/main/london_1800_1875_v0.5/fellowcitizens.png?raw=true)
+![TimeCapsuleLLM Sample Output](https://github.com/haykgrigo3/TimeCapsuleLLM/blob/main/london_1800_1875_v0.5/fellowcitizens.png?raw=true)
### **v1**
@@ -108,9 +108,9 @@ Charles DarwinECCEMACY. Sir, — The following case is interesting to me : — I
Trained using a 90GB (112GB tokenized) dataset
Trained to 182K steps
-![TimeLockLLM Sample Output](https://github.com/haykgrigo3/TimeCapsuleLLM/blob/main/london_1800_1875_v2/v2output1.png)
-
+![TimeCapsuleLLM Sample Output](https://github.com/haykgrigo3/TimeCapsuleLLM/blob/main/london_1800_1875_v2/v2output1.png)
+![TimeCapsuleLLM Sample Output](https://github.com/haykgrigo3/TimeCapsuleLLM/blob/main/london_1800_1875_v2/v2output2.png)
## Datasets
@@ -221,5 +221,6 @@ GPU: A100 SXM rented
+