63 lines · 1.3 KBRaw
.gitignore | 3 +++
.trail.yaml | 4 ++++
README.md | 3 +++
data.py | 10 ++++++++++
train.py | 6 ++++++
5 files changed, 26 insertions(+)
diff --git a/.gitignore b/.gitignore
new file mode 100644
index 0000000..7a1510d
--- /dev/null
+++ b/.gitignore
@@ -0,0 +1,3 @@
+logs/
+runs/
+reports/
diff --git a/.trail.yaml b/.trail.yaml
new file mode 100644
index 0000000..3a0b186
--- /dev/null
+++ b/.trail.yaml
@@ -0,0 +1,4 @@
+store: supabase://trail/main
+repo: ml-pipeline
+harnesses: [claude, codex]
+artifacts: ["logs/", "runs/**/metrics.json"]
diff --git a/README.md b/README.md
new file mode 100644
index 0000000..39050fc
--- /dev/null
+++ b/README.md
@@ -0,0 +1,3 @@
+# ml-pipeline
+
+Fine-tuning for the support assistant. LoRA on a 7B base model.
diff --git a/data.py b/data.py
new file mode 100644
index 0000000..92f70f8
--- /dev/null
+++ b/data.py
@@ -0,0 +1,10 @@
+import json
+
+
+def split(prompts, eval_frac=0.05):
+ seen = set()
+ for p in prompts:
+ key = p
+ if key in seen:
+ continue
+ seen.add(key)
diff --git a/train.py b/train.py
new file mode 100644
index 0000000..9bec56e
--- /dev/null
+++ b/train.py
@@ -0,0 +1,6 @@
+MODEL = "llama-7b"
+LORA_RANK = 64
+LR = 3e-4
+WARMUP_STEPS = 100
+GRAD_CLIP = None
+EARLY_STOP_PATIENCE = None