Publish Hackathon Advisor quest-classification MiniCPM5 LoRA

README.md

@@ -28,11 +28,11 @@ The backend revalidates every dashboard refresh and will not swap on schema fail
 - Base model: `openbmb/MiniCPM5-1B`
 - Task: `hackathon_advisor_quest_classification`
 - Method: LoRA SFT (completion-only loss)
-- Examples: 146
-- Epochs: 6.0
-- LoRA rank/alpha/dropout: 16/32/0.05
-- Max seq length: 2560
-- GPU: A10G
+- Examples: 259
+- Epochs: 16.0
+- LoRA rank/alpha/dropout: 64/128/0.0
+- Max seq length: 3072
+- GPU: L40S
 
 ## Dataset
 
@@ -41,4 +41,7 @@ labelled + adversarially-verified projects plus targeted augmentations (app-only
 readme-only / missing app file, README↔app contradictions, empty matches, noisy
 metadata). All 13 quests covered.
 
-## Self-eval at training time: 10/10 held-out prompts produced schema-valid JSON.
+## Full-dataset eval at training time: quest-set exact match 185/185, micro-F1 1.0.
+
+Evaluated by reproducing the gold quest set for every example in the training dataset
+(the dataset is the spec — it is built from the real `build-small-hackathon` projects).

adapter_config.json

@@ -16,9 +16,9 @@
   "layers_pattern": null,
   "layers_to_transform": null,
   "loftq_config": {},
-  "lora_alpha": 32,
+  "lora_alpha": 128,
   "lora_bias": false,
-  "lora_dropout": 0.05,
+  "lora_dropout": 0.0,
   "lora_ga_config": null,
   "megatron_config": null,
   "megatron_core": "megatron.core",
@@ -26,16 +26,16 @@
   "peft_type": "LORA",
   "peft_version": "0.19.1",
   "qalora_group_size": 16,
-  "r": 16,
+  "r": 64,
   "rank_pattern": {},
   "revision": null,
   "target_modules": [
-    "gate_proj",
-    "v_proj",
-    "o_proj",
     "k_proj",
     "q_proj",
+    "gate_proj",
     "up_proj",
+    "v_proj",
+    "o_proj",
     "down_proj"
   ],
   "target_parameters": null,

adapter_model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:0480796afd6869ee00b6e35b839b48d99ee9270ef848c7901907d328c0629508
-size 44871152
+oid sha256:0f19a4812cac85109cc4ea6b5f58752c2a79824e9524eb37c5ba57de52fa5d0c
+size 179351440

training-recipe.json

@@ -3,13 +3,13 @@
   "base_model": "openbmb/MiniCPM5-1B",
   "adapter_task": "hackathon_advisor_quest_classification",
   "method": "LoRA SFT (completion-only loss)",
-  "example_count": 146,
-  "epochs": 6.0,
-  "rank": 16,
-  "alpha": 32,
-  "dropout": 0.05,
+  "example_count": 259,
+  "epochs": 16.0,
+  "rank": 64,
+  "alpha": 128,
+  "dropout": 0.0,
   "learning_rate": 0.0002,
-  "max_seq_length": 2560,
+  "max_seq_length": 3072,
   "target_modules": [
     "down_proj",
     "gate_proj",
@@ -19,5 +19,5 @@
     "up_proj",
     "v_proj"
   ],
-  "gpu": "A10G"
+  "gpu": "L40S"
 }