Upload 163 files

hugging/td_fuse/selfimprove.py

@@ -787,14 +787,17 @@ def improve_with_lora(model_path: str, target_layers: List[int],
                 if len(train_texts) >= total:
                     break
         print(f"    Loaded {len(train_texts)} targeted samples")
-        if len(train_texts) < total * 0.5:
-            print(f"    WARNING: Only got {len(train_texts)}/{total} targeted samples.")
-            print(f"    Supplementing with generic data to reach {total}...")
+        if len(train_texts) < 200:
+            print(f"    WARNING: Only got {len(train_texts)} targeted samples (very low).")
+            print(f"    Supplementing with generic data...")
             # Fall through to hardcoded data loading below to fill the gap
         else:
-            # Got enough targeted data — skip hardcoded loading
-            # Jump past the hardcoded data section
-            pass  # train_texts is already populated
+            # Got targeted data — DON'T dilute with easy generic data
+            # Even 500 hard targeted samples > 8000 easy samples the model already knows
+            # Training loss was 0.0000 when we mixed targeted + generic (model knew the generic data)
+            print(f"    Using {len(train_texts)} targeted samples ONLY (no generic filler)")
+            print(f"    Reason: generic data causes loss=0 because model already knows it")
+            total = len(train_texts)  # Adjust total so we don't try to supplement
 
     # Only load hardcoded datasets if we don't have enough targeted data
     # Only load generic datasets if targeted data wasn't enough

hugging/td_fuse/td_data_finder.py

@@ -377,44 +377,53 @@ def _format_hf_item(item: Dict, category: str, source: str) -> Optional[Formatte
 # Verified dataset names on HuggingFace (March 2026)
 # Each tuple: (dataset_name, config_or_subset)
 # If a dataset fails to load, it moves to the next one silently
+# Each tuple: (dataset_name, config_or_subset, split_override)
+# split_override: if None, uses "train". Set to specific split name if needed.
+# If a dataset fails to load, it moves to the next one silently.
 FALLBACK_DATASETS = {
     "math": [
-        ("openai/gsm8k", "main"),          # 8.8K train, question/answer with ####
-        ("lighteval/MATH", "all"),           # 12.5K, competition math
-        ("microsoft/orca-math-word-problems-200k", None),  # 200K word problems
-        ("AI-MO/NuminaMath-CoT", None),     # 860K chain-of-thought math
-        ("TIGER-Lab/MathInstruct", None),   # 262K math instruction pairs
+        ("openai/gsm8k", "main", None),                    # 8.8K, question/answer with ####
+        ("lighteval/MATH", "all", None),                     # 12.5K, competition math (HARD)
+        ("AI-MO/NuminaMath-CoT", None, None),               # 860K chain-of-thought math
+        ("TIGER-Lab/MathInstruct", None, None),             # 262K math instruction pairs
+        ("microsoft/orca-math-word-problems-200k", None, None),  # 200K word problems
     ],
     "code": [
-        ("openai/openai_humaneval", None),   # 164 problems — small but high quality
-        ("google-research-datasets/mbpp", None),  # FIX: full name (short "mbpp" fails)
-        ("codeparrot/apps", "introductory"), # Competitive programming
-        ("m-a-p/CodeFeedback-Filtered-Instruction", None),  # 157K code instructions
+        ("m-a-p/CodeFeedback-Filtered-Instruction", None, None),  # 157K code instructions — WORKS
+        ("openai/openai_humaneval", None, "test"),           # 164 problems — FIX: split is "test"
+        ("google-research-datasets/mbpp", None, "test"),     # FIX: split is "test"
+        ("bigcode/starcoderdata", None, None),               # StarCoder training data
+        ("sahil2801/CodeAlpaca-20k", None, None),            # 20K code instruction pairs
     ],
     "reasoning": [
-        ("allenai/ai2_arc", "ARC-Challenge"),  # 2.6K challenge set
-        ("Rowan/hellaswag", None),              # 40K commonsense
-        ("allenai/openbookqa", "main"),         # FIX: replaced openai_prm800k (not a public dataset)
-        ("allenai/winogrande", "winogrande_xl"),  # 40K commonsense reasoning
-        ("tasksource/bigbench", None),          # Diverse reasoning tasks
+        ("allenai/ai2_arc", "ARC-Challenge", None),         # 2.6K challenge set
+        ("Rowan/hellaswag", None, None),                     # 40K commonsense
+        ("allenai/openbookqa", "main", None),                # Science reasoning
+        ("allenai/winogrande", "winogrande_xl", None),       # 40K commonsense reasoning
+        ("TIGER-Lab/MathInstruct", None, None),             # Also good for reasoning chains
     ],
     "creativity": [
-        ("teknium/OpenHermes-2.5", None),   # 1M instruction pairs — diverse, includes creative
-        ("Open-Orca/OpenOrca", None),        # 4.2M conversations
-        ("tatsu-lab/alpaca", None),           # 52K instructions — many creative tasks
-        ("WizardLMTeam/WizardLM_evol_instruct_V2_196k", None),  # 196K evolved instructions
+        ("teknium/OpenHermes-2.5", None, None),             # 1M instruction pairs — diverse
+        ("WizardLMTeam/WizardLM_evol_instruct_V2_196k", None, None),  # 196K evolved (HARDER)
+        ("tatsu-lab/alpaca", None, None),                    # 52K instructions
+        ("Open-Orca/OpenOrca", None, None),                  # 4.2M conversations
     ],
     "knowledge": [
-        ("TIGER-Lab/MMLU-Pro", None),        # MMLU harder version
-        ("cais/mmlu", "all"),                 # Classic MMLU
-        ("truthfulqa/truthful_qa", "generation"),  # 817 truthfulness questions
-        ("sail/regmix-data-sample", None),   # Diverse knowledge mix
+        ("cais/mmlu", "all", "test"),                        # FIX: split is "test" not "train"
+        ("cais/mmlu", "all", "validation"),                  # Also try validation split
+        ("truthfulqa/truthful_qa", "generation", "validation"),  # FIX: split is "validation"
+        ("lukaemon/mmlu", "anatomy", "test"),                # Individual MMLU subject
+        ("lukaemon/mmlu", "world_religions", "test"),        # More knowledge topics
+        ("lukaemon/mmlu", "geography", "test"),              # Geography specifically (weak spot!)
+        ("lukaemon/mmlu", "astronomy", "test"),              # Science knowledge
+        ("lukaemon/mmlu", "global_facts", "test"),           # General knowledge
+        ("Stevross/Erehwon-20k", None, None),               # 20K diverse knowledge QA
     ],
     "instruction_following": [
-        ("HuggingFaceH4/ultrafeedback_binarized", None),  # 61K preference pairs
-        ("tatsu-lab/alpaca", None),           # 52K instructions
-        ("Intel/orca_dpo_pairs", None),       # 12K DPO pairs — strong instruction following
-        ("argilla/ultrafeedback-binarized-preferences", None),  # Alternative format
+        ("tatsu-lab/alpaca", None, None),                    # 52K instructions — WORKS
+        ("HuggingFaceH4/ultrafeedback_binarized", None, "train_prefs"),  # FIX: correct split name
+        ("Intel/orca_dpo_pairs", None, None),                # 12K DPO pairs
+        ("argilla/ultrafeedback-binarized-preferences", None, None),
     ],
 }
 
@@ -474,16 +483,22 @@ def find_training_data(
             except ImportError:
                 print("    Warning: datasets library not installed — cannot load datasets")
                 fallbacks = []  # Skip the fallback loop
-            for ds_name, config in fallbacks:
+            for fallback_entry in fallbacks:
                 if len(category_samples) >= needed:
                     break
 
+                # Unpack: (name, config, split_override) — split_override can be None
+                ds_name = fallback_entry[0]
+                config = fallback_entry[1]
+                split_override = fallback_entry[2] if len(fallback_entry) > 2 else None
+                use_split = split_override or "train"
+
                 remaining = needed - len(category_samples)
                 try:
                     if config:
-                        ds = load_dataset(ds_name, config, streaming=True, split="train")
+                        ds = load_dataset(ds_name, config, streaming=True, split=use_split)
                     else:
-                        ds = load_dataset(ds_name, streaming=True, split="train")
+                        ds = load_dataset(ds_name, streaming=True, split=use_split)
 
                     count = 0
                     stall = 0
@@ -531,7 +546,7 @@ def find_training_data(
                     break
 
                 # Skip datasets we already used as fallbacks
-                if any(result.name == fb[0] for fb in fallbacks):
+                if any(result.name == fb_entry[0] for fb_entry in fallbacks):
                     continue
 
                 remaining = needed - len(category_samples)