document temperature and system prompt sensitivity

contrib/fine-tuning/Modelfile

@@ -1,11 +1,9 @@
 FROM ./cdx1-gguf-q8_0.gguf
 
-PARAMETER num_ctx 16384
+PARAMETER num_ctx 16000
 PARAMETER temperature 0.05
-PARAMETER top_k 10
-PARAMETER top_p 0.5
 
-SYSTEM """You are cdxgen, an expert in CycloneDX and xBOM."""
+SYSTEM """You are a helpful assistant."""
 
 LICENSE """
 apache-2.0

contrib/fine-tuning/README.md

@@ -30,50 +30,65 @@ bash fine-tune-mlx.sh
 cp -rf prabhuat ~/.lmstudio/models/
 lms ls
 lms server status
-lms load CycloneDX/cdx1-mlx --exact --gpu max --identifier cdx1-test --context-length 8192
+lms load CycloneDX/cdx1-mlx-8bit --exact --gpu max --identifier cdx1-test --context-length 16000
 ```
 
 System prompt:
 
 ```text
-You are cdxgen, an expert in CycloneDX and xBOM.
+You are a helpful assistant.
 ```
 
-### gguf testing with ollama
+### Testing with ollama
 
-Use the generated `Modelfile` inside `CycloneDX/cdx1-gguf` to test cdx1 with ollama.
+Create a Modelfile as shown:
+
+```text
+FROM hf.co/CycloneDX/cdx1-gguf-Q8_0-GGUF
+
+PARAMETER num_ctx 16000
+PARAMETER temperature 0.05
+
+SYSTEM """You are a helpful assistant."""
+```
 
 ```shell
-cd CycloneDX/cdx1-gguf
-ollama create cdx1-gguf -f ./Modelfile
+ollama create cdx1 -f ./Modelfile
 ```
 
 ```text
-ollama show cdx1-gguf
-  Model
+ollama show cdx1
+Model
     architecture        llama
     parameters          14.7B
     context length      16384
     embedding length    5120
-    quantization        F16
+    quantization        Q8_0
 
   Parameters
-    num_ctx        16384
+    num_ctx        16000
+    stop           "<|im_start|>"
+    stop           "<|im_sep|>"
+    stop           "<|im_end|>"
+    stop           "<|im_start|>"
+    stop           "<|im_sep|>"
+    stop           "<|im_end|>"
+    stop           "<|im_start|>"
+    stop           "<|im_sep|>"
+    stop           "<|im_end|>"
+    stop           "<|im_start|>user<|im_sep|>"
     temperature    0.05
-    top_k          10
-    top_p          0.5
 
   System
-    You are cdxgen, an expert in CycloneDX and xBOM.
-
-  License
-    apache-2.0
+    You are a helpful assistant.
 ```
 
 ```shell
-ollama run cdx1-gguf "Tell me about cdxgen"
+ollama run cdx1 "Tell me about cdxgen"
 ```
 
+Use `hf.co/CycloneDX/cdx1-gguf-BF16-GGUF` for higher precision needs.
+
 ### Validating jsonlines files
 
 ```shell

contrib/fine-tuning/convert-gguf.sh

@@ -13,17 +13,25 @@ TOOL_BASE_MODEL=cdx1
 LLAMA_CPP_PATH=/Volumes/Work/sandbox/llama.cpp
 cd $LLAMA_CPP_PATH
 CDXGEN_FT_PATH=/Volumes/Work/CycloneDX/cdxgen/contrib/fine-tuning
+
 GGUF_MODEL_Q8_0_NAME=${HF_ORG}/${TOOL_BASE_MODEL}-gguf-Q8_0-GGUF
 GGUF_MODEL_Q8_0_PATH=${CDXGEN_FT_PATH}/${HF_ORG}/${TOOL_BASE_MODEL}-gguf-Q8_0-GGUF
 FUSED_MODEL=${CDXGEN_FT_PATH}/${HF_ORG}/${TOOL_BASE_MODEL}-${TUNING_TOOL}
 
 rm -rf ${GGUF_MODEL_Q8_0_PATH}
 mkdir -p ${GGUF_MODEL_Q8_0_PATH}
 python convert_hf_to_gguf.py --outtype q8_0 --outfile ${CDXGEN_FT_PATH}/${HF_ORG}/${TOOL_BASE_MODEL}-gguf-Q8_0-GGUF/${TOOL_BASE_MODEL}-gguf-q8_0.gguf --model-name ${GGUF_MODEL_Q8_0_NAME} ${FUSED_MODEL}
-
 cp ${CDXGEN_FT_PATH}/Modelfile ${GGUF_MODEL_Q8_0_PATH}/
 # cp ${FUSED_MODEL}/*.json ${FUSED_MODEL}/merges.txt ${GGUF_MODEL_Q8_0_PATH}/
 
+GGUF_MODEL_BF16_0_NAME=${HF_ORG}/${TOOL_BASE_MODEL}-gguf-BF16-GGUF
+GGUF_MODEL_BF16_0_PATH=${CDXGEN_FT_PATH}/${HF_ORG}/${TOOL_BASE_MODEL}-gguf-BF16-GGUF
+rm -rf ${GGUF_MODEL_BF16_0_PATH}
+mkdir -p ${GGUF_MODEL_BF16_0_PATH}
+python convert_hf_to_gguf.py --outtype bf16 --outfile ${CDXGEN_FT_PATH}/${HF_ORG}/${TOOL_BASE_MODEL}-gguf-BF16-GGUF/${TOOL_BASE_MODEL}-gguf-bf16.gguf --model-name ${GGUF_MODEL_BF16_0_NAME} ${FUSED_MODEL}
+cp ${CDXGEN_FT_PATH}/Modelfile ${GGUF_MODEL_BF16_0_PATH}/
+sed -i '' 's|./cdx1-gguf-q8_0.gguf|./cdx1-gguf-bf16.gguf|g' ${GGUF_MODEL_BF16_0_PATH}/Modelfile
+
 ### Testing with ollama
 # cd ${GGUF_MODEL_Q8_0_PATH}
 # ollama create cdx1-gguf -f Modelfile
@@ -33,3 +41,4 @@ cp ${CDXGEN_FT_PATH}/Modelfile ${GGUF_MODEL_Q8_0_PATH}/
 export HF_HUB_ENABLE_HF_TRANSFER=0
 huggingface-cli whoami
 huggingface-cli upload --quiet --repo-type model ${GGUF_MODEL_Q8_0_NAME} ${GGUF_MODEL_Q8_0_PATH} .
+huggingface-cli upload --quiet --repo-type model ${GGUF_MODEL_BF16_0_NAME} ${GGUF_MODEL_BF16_0_PATH} .

contrib/xBOMEval/README.md

@@ -0,0 +1,2 @@
+# Introduction
+