Update app.py

app.py

@@ -586,13 +586,31 @@ def wrapper_specious(token, method, base, norm, i8, t, filt_w, m1, w1, f1, m2, w
 
     yield from run_mergekit_logic(config, token, out, priv, shard, prec, tok_src, chat_t, program="mergekit-yaml")
 
-def wrapper_moer(token, base, experts, gate, dtype, out, priv, shard, prec, tok_src, chat_t):
-    formatted = [{"source_model": e.strip(), "positive_prompts": ["chat", "assist"]} for e in experts.split('\n') if e.strip()]
+def wrapper_moer(token, base, expert1, prompt1, expert2, prompt2, expert3, prompt3, expert4, prompt4, expert5, prompt5, gate, dtype, out, priv, shard, prec, tok_src, chat_t):
+    experts = []
+    for exp, pmt in [
+        (expert1, prompt1), (expert2, prompt2), (expert3, prompt3),
+        (expert4, prompt4), (expert5, prompt5)
+    ]:
+        if exp.strip():
+            expert_entry = {"source_model": exp.strip()}
+            # Parse prompts (comma-separated)
+            if pmt.strip():
+                prompts = [p.strip() for p in pmt.split(',') if p.strip()]
+                expert_entry["positive_prompts"] = prompts
+            else:
+                expert_entry["positive_prompts"] = [""]
+            experts.append(expert_entry)
+    
+    if len(experts) < 2:
+        return "Error: At least 2 experts required"
+    
+    # Build config for MoE
     config = {
-        "base_model": base.strip() if base.strip() else formatted[0]["source_model"],
+        "base_model": base.strip(),
         "gate_mode": gate,
         "dtype": dtype, 
-        "experts": formatted
+        "experts": experts
     }
     # Uses mergekit-moe CLI
     yield from run_mergekit_logic(config, token, out, priv, shard, prec, tok_src, chat_t, program="mergekit-moe")
@@ -700,21 +718,23 @@ with gr.Blocks() as demo:
 
         # --- TAB 5 ---
         with gr.Tab("Amphinterpolative"):
-            gr.Markdown("### Spherical Interpolation Family")
+            gr.Markdown("### Spherical Interpolation Methods Family: slerp, nuslerp, multislerp, karcher")
             t5_token = gr.Textbox(label="HF Token", type="password")
-            t5_method = gr.Dropdown(["slerp", "nuslerp", "multislerp", "karcher"], value="slerp", label="Method")
+            t5_method = gr.Dropdown(["slerp", "nuslerp", "multislerp", "karcher"], value="slerp", label="Merge Method")
+            gr.Markdown("See [MergeKit Merge Method Docs](https://github.com/arcee-ai/mergekit/blob/main/docs/merge_methods.md) for more info.")
             with gr.Row():
-                t5_shard = gr.Slider(label="Max Shard Size (GB)", value=5.0, minimum=1.0, maximum=20.0)
+                t5_shard = gr.Slider(label="Max Shard Size (GB)", value=5.0, minimum=0.5, maximum=20.0)
                 t5_prec = gr.Dropdown(["float16", "bfloat16", "float32"], value="bfloat16", label="Output Precision")
                 t5_tok = gr.Dropdown(["base", "union", "model:path"], value="base", label="Tokenizer Source")
-                t5_chat = gr.Textbox(label="Chat Template (write-in, default: auto)", placeholder="auto")
+                t5_chat = gr.Textbox(label="Chat Template (default: auto)", placeholder="auto")
+            gr.Markdown("Built-in Chat Templates: alpaca, chatml, llama3, mistral, exaone, auto")
             with gr.Row():
                 t5_base = gr.Textbox(label="Base Model")
                 t5_t = gr.Slider(0, 1, 0.5, label="t")
             with gr.Row():
-                t5_norm = gr.Checkbox(label="Normalize", value=True); t5_i8 = gr.Checkbox(label="Int8 Mask", value=False); t5_flat = gr.Checkbox(label="NuSlerp Flatten", value=False); t5_row = gr.Checkbox(label="NuSlerp Row Wise", value=False)
+                t5_norm = gr.Checkbox(label="Normalize Weights", value=True); t5_i8 = gr.Checkbox(label="Int8 Mask", value=False); t5_flat = gr.Checkbox(label="Flatten Tensors (NuSlerp)", value=False); t5_row = gr.Checkbox(label="Row Wise (NuSlerp)", value=False)
             with gr.Row():
-                t5_eps = gr.Textbox(label="eps", value="1e-8"); t5_iter = gr.Number(label="max_iter", value=10); t5_tol = gr.Textbox(label="tol", value="1e-5")
+                t5_eps = gr.Textbox(label="eps (Stabilization Constant) (MultiSlerp)", value="1e-8"); t5_iter = gr.Number(label="Max Iterations (Karcher)", value=10); t5_tol = gr.Textbox(label="tol (Convergence Tolerance) (Karcher)", value="1e-5")
             m1, w1 = gr.Textbox(label="Model 1"), gr.Textbox(label="Weight 1", value="1.0"); m2, w2 = gr.Textbox(label="Model 2"), gr.Textbox(label="Weight 2", value="1.0")
             with gr.Accordion("More", open=False):
                 m3, w3 = gr.Textbox(label="Model 3"), gr.Textbox(label="Weight 3", value="1.0"); m4, w4 = gr.Textbox(label="Model 4"), gr.Textbox(label="Weight 4", value="1.0"); m5, w5 = gr.Textbox(label="Model 5"), gr.Textbox(label="Weight 5", value="1.0")
@@ -725,19 +745,19 @@ with gr.Blocks() as demo:
 
         # --- TAB 6 ---
         with gr.Tab("Stir/Tie Bases"):
-            gr.Markdown("### Task Vector Family")
-            t6_token = gr.Textbox(label="Token", type="password")
-            t6_method = gr.Dropdown(["task_arithmetic", "ties", "dare_ties", "dare_linear", "della", "della_linear", "breadcrumbs", "breadcrumbs_ties", "sce"], value="ties", label="Method")
+            gr.Markdown("### Task Vector Methods Family: task_arithmetic, ties, dare_ties, dare_linear, della, della_linear, breadcrumbs, breadcrumbs_ties, sce")            t6_token = gr.Textbox(label="Token", type="password")
+            t6_method = gr.Dropdown(["task_arithmetic", "ties", "dare_ties", "dare_linear", "della", "della_linear", "breadcrumbs", "breadcrumbs_ties", "sce"], value="ties", label="Merge Method")
+            gr.Markdown("See [MergeKit Merge Method Docs](https://github.com/arcee-ai/mergekit/blob/main/docs/merge_methods.md) for more info.")
             with gr.Row():
-                t6_shard = gr.Slider(label="Max Shard Size (GB)", value=5.0, minimum=1.0, maximum=20.0); t6_prec = gr.Dropdown(["float16", "bfloat16", "float32"], value="bfloat16", label="Output Precision"); t6_tok = gr.Dropdown(["base", "union", "model:path"], value="base", label="Tokenizer Source"); t6_chat = gr.Textbox(label="Chat Template", placeholder="auto")
-            t6_base = gr.Textbox(label="Base Model")
+                t6_shard = gr.Slider(label="Max Shard Size (GB)", value=5.0, minimum=0.5, maximum=20.0); t6_prec = gr.Dropdown(["float16", "bfloat16", "float32"], value="bfloat16", label="Output Precision"); t6_tok = gr.Dropdown(["base", "union", "model:path"], value="base", label="Tokenizer Source"); t6_chat = gr.Textbox(label="Chat Template", placeholder="auto")
+            t6_base = gr.Textbox(label="Base Model (required)")
             with gr.Row():
-                t6_norm = gr.Checkbox(label="Normalize", value=True); t6_i8 = gr.Checkbox(label="Int8 Mask", value=False); t6_resc = gr.Checkbox(label="Rescale", value=True); t6_lamb = gr.Number(label="Lambda", value=1.0); t6_topk = gr.Slider(0, 1, 1.0, label="Select TopK")
+                t6_norm = gr.Checkbox(label="Normalize Weights", value=True); t6_i8 = gr.Checkbox(label="Int8 Mask", value=False); t6_resc = gr.Checkbox(label="Rescale (Dare_Linear)", value=True); t6_lamb = gr.Number(label="Lambda", value=1.0); t6_topk = gr.Slider(0, 1, 1.0, label="Select TopK (SCE)")
             m1_6, w1_6 = gr.Textbox(label="Model 1"), gr.Textbox(label="Weight 1", value="1.0"); d1_6, g1_6, e1_6 = gr.Textbox(label="Density", value="1.0"), gr.Number(label="Gamma", value=0.01), gr.Number(label="Epsilon", value=0.15)
             with gr.Accordion("More", open=False):
-                m2_6, w2_6 = gr.Textbox(label="Model 2"), gr.Textbox(label="Weight 2", value="1.0"); d2_6, g2_6, e2_6 = gr.Textbox(label="Density", value="1.0"), gr.Number(label="Gamma", value=0.01), gr.Number(label="Epsilon", value=0.15)
-                m3_6, w3_6 = gr.Textbox(label="Model 3"), gr.Textbox(label="Weight 3", value="1.0"); d3_6, g3_6, e3_6 = gr.Textbox(label="Density", value="1.0"), gr.Number(label="Gamma", value=0.01), gr.Number(label="Epsilon", value=0.15)
-                m4_6, w4_6 = gr.Textbox(label="Model 4"), gr.Textbox(label="Weight 4", value="1.0"); d4_6, g4_6, e4_6 = gr.Textbox(label="Density", value="1.0"), gr.Number(label="Gamma", value=0.01), gr.Number(label="Epsilon", value=0.15)
+                m2_6, w2_6 = gr.Textbox(label="Model 2"), gr.Textbox(label="Weight 2", value="1.0"); d2_6, g2_6, e2_6 = gr.Textbox(label="Density (DARE/TIES)", value="1.0"), gr.Number(label="Gamma (breadcrumbs)", value=0.01), gr.Number(label="Epsilon (DELLA)", value=0.15)
+                m3_6, w3_6 = gr.Textbox(label="Model 3"), gr.Textbox(label="Weight 3", value="1.0"); d3_6, g3_6, e3_6 = gr.Textbox(label="Density (DARE/TIES)", value="1.0"), gr.Number(label="Gamma (breadcrumbs)", value=0.01), gr.Number(label="Epsilon (DELLA)", value=0.15)
+                m4_6, w4_6 = gr.Textbox(label="Model 4"), gr.Textbox(label="Weight 4", value="1.0"); d4_6, g4_6, e4_6 = gr.Textbox(label="Density (DARE/TIES)", value="1.0"), gr.Number(label="Gamma (breadcrumbs)", value=0.01), gr.Number(label="Epsilon (DELLA)", value=0.15)
             t6_out = gr.Textbox(label="Output Repo"); t6_priv = gr.Checkbox(label="Private", value=True)
             t6_btn = gr.Button("Execute")
             t6_res = gr.Textbox(label="Result", lines=10)
@@ -745,15 +765,21 @@ with gr.Blocks() as demo:
 
         # --- TAB 7 ---
         with gr.Tab("Specious"):
-            gr.Markdown("### Specialized Methods")
+            gr.Markdown("### Specialized Methods: model_stock, nearswap, arcee_fusion, passthrough")
             t7_token = gr.Textbox(label="Token", type="password")
-            t7_method = gr.Dropdown(["model_stock", "nearswap", "arcee_fusion", "passthrough", "linear"], value="model_stock", label="Method")
+            t7_method = gr.Dropdown(["model_stock", "nearswap", "arcee_fusion", "passthrough", "linear"], value="model_stock", label="Merge Method")
+            gr.Markdown("See [MergeKit Merge Method Docs](https://github.com/arcee-ai/mergekit/blob/main/docs/merge_methods.md) for more info.")
             with gr.Row():
-                t7_shard = gr.Slider(label="Max Shard Size (GB)", value=5.0, minimum=1.0, maximum=20.0); t7_prec = gr.Dropdown(["float16", "bfloat16", "float32"], value="bfloat16", label="Output Precision"); t7_tok = gr.Dropdown(["base", "union", "model:path"], value="base", label="Tokenizer Source"); t7_chat = gr.Textbox(label="Chat Template", placeholder="auto")
-            t7_base = gr.Textbox(label="Base Model")
+                t7_shard = gr.Slider(label="Max Shard Size (GB)", value=5.0, minimum=0.5, maximum=20.0); t7_prec = gr.Dropdown(["float16", "bfloat16", "float32"], value="bfloat16", label="Output Precision"); t7_tok = gr.Dropdown(["base", "union", "model:path"], value="base", label="Tokenizer Source"); t7_chat = gr.Textbox(label="Chat Template", placeholder="auto")
+            
+            t7_base = gr.Textbox(label="Base Model (required for nearswap/arcee_fusion/model_stock)", placeholder="org/base-model")
+            
+            gr.Markdown("#### Models")
+            gr.Markdown("**passthrough:** 1 model | **nearswap/arcee_fusion:** 2 models | **model_stock:** 3+ models")
+            
             with gr.Row():
-                t7_norm = gr.Checkbox(label="Normalize", value=True); t7_i8 = gr.Checkbox(label="Int8 Mask", value=False); t7_t = gr.Slider(0, 1, 0.5, label="t"); t7_filt_w = gr.Checkbox(label="Filter Wise", value=False)
-            m1_7, w1_7, f1_7 = gr.Textbox(label="Model 1"), gr.Textbox(label="Weight 1", value="1.0"), gr.Textbox(label="Filter (Passthrough)")
+                t7_norm = gr.Checkbox(label="Normalize", value=True); t7_i8 = gr.Checkbox(label="Int8 Mask", value=False); t7_t = gr.Slider(0, 1, 0.5, label="t (Interpolation Ratio, for Nearswap)"); t7_filt_w = gr.Checkbox(label="Filter Wise (for Model_Stock)", value=False)
+            m1_7, w1_7, f1_7 = gr.Textbox(label="Model 1"), gr.Textbox(label="Weight 1", value="1.0"), gr.Textbox(label="Filter Model Component")
             m2_7, w2_7 = gr.Textbox(label="Model 2"), gr.Textbox(label="Weight 2", value="1.0")
             with gr.Accordion("More", open=False):
                 m3_7, w3_7 = gr.Textbox(label="Model 3"), gr.Textbox(label="Weight 3", value="1.0"); m4_7, w4_7 = gr.Textbox(label="Model 4"), gr.Textbox(label="Weight 4", value="1.0"); m5_7, w5_7 = gr.Textbox(label="Model 5"), gr.Textbox(label="Weight 5", value="1.0")
@@ -765,22 +791,48 @@ with gr.Blocks() as demo:
         # --- TAB 8 (MoEr) ---
         with gr.Tab("MoEr"):
             gr.Markdown("### Mixture of Experts")
+            gr.Markdown("See [MergeKit MoE doc for more info](https://github.com/arcee-ai/mergekit/blob/main/docs/moe.md) for more info.")
+
             t8_token = gr.Textbox(label="Token", type="password")
             with gr.Row():
-                t8_shard = gr.Slider(label="Max Shard Size (GB)", value=5.0, minimum=1.0, maximum=20.0); t8_prec = gr.Dropdown(["float16", "bfloat16", "float32"], value="bfloat16", label="Output Precision"); t8_tok = gr.Dropdown(["base", "union", "model:path"], value="base", label="Tokenizer Source"); t8_chat = gr.Textbox(label="Chat Template", placeholder="auto")
+                t8_shard = gr.Slider(label="Max Shard Size (GB)", value=5.0, minimum=0.5, maximum=20.0); t8_prec = gr.Dropdown(["float16", "bfloat16", "float32"], value="bfloat16", label="Output Precision"); t8_tok = gr.Dropdown(["base", "union", "model:path"], value="base", label="Tokenizer Source"); t8_chat = gr.Textbox(label="Chat Template", placeholder="auto")
             t8_base = gr.Textbox(label="Base Model"); t8_experts = gr.TextArea(label="Experts List"); t8_gate = gr.Dropdown(["cheap_embed", "random", "hidden"], value="cheap_embed", label="Gate Mode"); t8_dtype = gr.Dropdown(["float16", "bfloat16"], value="bfloat16", label="Internal Dtype")
+            gr.Markdown("#### Experts (at least 2 required)")
+            gr.Markdown("Prompts are comma-separated descriptors for each expert")
+            
+            with gr.Row():
+                t8_expert1 = gr.Textbox(label="Expert 1", placeholder="org/expert1")
+                t8_prompt1 = gr.Textbox(label="Positive Prompts", placeholder="math, reasoning, logic")
+            
+            with gr.Row():
+                t8_expert2 = gr.Textbox(label="Expert 2", placeholder="org/expert2")
+                t8_prompt2 = gr.Textbox(label="Positive Prompts", placeholder="creative, writing, storytelling")
+            
+            with gr.Row():
+                t8_expert3 = gr.Textbox(label="Expert 3 (optional)", placeholder="org/expert3")
+                t8_prompt3 = gr.Textbox(label="Positive Prompts", placeholder="code, programming")
+            
+            with gr.Row():
+                t8_expert4 = gr.Textbox(label="Expert 4 (optional)", placeholder="org/expert4")
+                t8_prompt4 = gr.Textbox(label="Positive Prompts", placeholder="")
+            
+            with gr.Row():
+                t8_expert5 = gr.Textbox(label="Expert 5 (optional)", placeholder="org/expert5")
+                t8_prompt5 = gr.Textbox(label="Positive Prompts", placeholder="")
             t8_out = gr.Textbox(label="Output Repo"); t8_priv = gr.Checkbox(label="Private", value=True)
             t8_btn = gr.Button("Build MoE")
             t8_res = gr.Textbox(label="Result", lines=10)
-            t8_btn.click(wrapper_moer, [t8_token, t8_base, t8_experts, t8_gate, t8_dtype, t8_out, t8_priv, t8_shard, t8_prec, t8_tok, t8_chat], t8_res)
+            t8_btn.click(wrapper_moer, [t8_token, t8_base, t8_expert1, t8_prompt1, t8_expert2, t8_prompt2, t8_expert3, t8_prompt3, t8_expert4, t8_prompt4, t8_expert5, t8_prompt5, t8_gate, t8_dtype, t8_out, t8_priv, t8_shard, t8_prec, t8_tok, t8_chat], t8_res)
 
         # --- TAB 9 (Rawer) ---
         with gr.Tab("Rawer"):
-            gr.Markdown("### Raw PyTorch / Non-Transformer")
+            gr.Markdown("### Raw PyTorch MergeKit / Non-pipeline-classed Models")
             t9_token = gr.Textbox(label="Token", type="password"); t9_models = gr.TextArea(label="Models (one per line)")
             with gr.Row():
-                t9_shard = gr.Slider(label="Max Shard Size (GB)", value=5.0, minimum=1.0, maximum=20.0); t9_prec = gr.Dropdown(["float16", "bfloat16", "float32"], value="bfloat16", label="Output Precision"); t9_tok = gr.Dropdown(["base", "union", "model:path"], value="base", label="Tokenizer Source"); t9_chat = gr.Textbox(label="Chat Template", placeholder="auto")
-            t9_method = gr.Dropdown(["linear", "passthrough"], value="linear", label="Method"); t9_dtype = gr.Dropdown(["float32", "float16", "bfloat16"], value="float32", label="Config Dtype")
+                t9_shard = gr.Slider(label="Max Shard Size (GB)", value=5.0, minimum=0.5, maximum=20.0); t9_prec = gr.Dropdown(["float16", "bfloat16", "float32"], value="bfloat16", label="Output Precision"); t9_tok = gr.Dropdown(["base", "union", "model:path"], value="base", label="Tokenizer Source"); t9_chat = gr.Textbox(label="Chat Template", placeholder="auto")
+            gr.Markdown("Built-in Chat Templates: alpaca, chatml, llama3, mistral, exaone, auto")
+            gr.Markdown("See [MergeKit Merge Method Docs](https://github.com/arcee-ai/mergekit/blob/main/docs/merge_methods.md) for more info.")
+            t9_method = gr.Dropdown(["linear", "passthrough"], value="linear", label="Merge Method"); t9_dtype = gr.Dropdown(["float32", "float16", "bfloat16"], value="float32", label="Config dtype")
             t9_out = gr.Textbox(label="Output Repo"); t9_priv = gr.Checkbox(label="Private", value=True)
             t9_btn = gr.Button("Merge Raw")
             t9_res = gr.Textbox(label="Result", lines=10)
@@ -788,11 +840,26 @@ with gr.Blocks() as demo:
 
         # --- TAB 10 ---
         with gr.Tab("Mario,DARE!"):
+            gr.Markdown("### Model-Agnostic DARE Implementation (Drop And REscale)")
+            gr.Markdown("From [sft-merger by Martyn Garcia](https://github.com/martyn)")
             t10_token = gr.Textbox(label="Token", type="password")
+            
+            gr.Markdown(
+                """
+                ### How DARE Works:
+                1. **Compute Delta**: Difference between fine-tuned and base weights
+                2. **Drop Elements**: Randomly mask out delta values based on mask rate
+                3. **Rescale**: Compensate for dropped elements by rescaling remaining values
+                4. **Apply**: Add scaled delta back to base model
+                
+                **Mask Rate**: 0.5 = drop 50% of delta values, 0.9 = drop 90% (more aggressive sparsification)
+                """
+            )
+            
             with gr.Row():
-                t10_base = gr.Textbox(label="Base Model"); t10_ft = gr.Textbox(label="Fine-Tuned Model")
+                t10_base = gr.Textbox(label="Base Model", placeholder="org/base-model"); t10_ft = gr.Textbox(label="Fine-Tuned Model", placeholder="org/fine-tuned-model")
             with gr.Row():
-                t10_ratio = gr.Slider(0, 5, 1.0, label="Ratio"); t10_mask = gr.Slider(0, 0.99, 0.5, label="Mask Rate")
+                t10_ratio = gr.Slider(value=1.0, minimum=0.0, maximum=2.0, step=0.1, label="Merge Ratio (delta weight)"); t10_mask = gr.Slider(value=0.5, minimum=0.0, maximum=0.99, step=0.01, label="Mask Rate (drop probability)")
             t10_out = gr.Textbox(label="Output Repo"); t10_priv = gr.Checkbox(label="Private", value=True)
             gr.Button("Run").click(task_dare_custom, [t10_token, t10_base, t10_ft, t10_ratio, t10_mask, t10_out, t10_priv], gr.Textbox(label="Result"))