Upload README.md with huggingface_hub

README.md

@@ -52,6 +52,7 @@ Each model answered the same calorie question on the same dishes
 | Qwen3.5-9B (Q8_0, thinking off) ‡ | 304 | 117.2 | 64.5 | 65.8% | +6.1 |
 | MiniCPM-V 4.6 (1.3B) | 304 | 163.3 | 97.0 | 52.6% | +75.5 |
 | Gemma-4-12B † | 100 | 176.8 | 90.5 | 58.0% | −43.3 |
+| MiniMax M3 (hosted, thinking) | 304 | 193.5 | 120.5 | 44.7% | +103.7 |
 | CalorieCLIP | 304 | 196.2 | 133.2 | 42.1% | +121.1 |
 
 † Gemma-4-12B was scored on a seeded 100-dish paired subset of the same 304 dishes (compute budget);
@@ -59,6 +60,33 @@ all parse rates were 100%.
 ‡ Qwen3.5-9B ran as an 8-bit GGUF (Q8_0) via llama.cpp with 4k context and thinking disabled.
 Notably it matches the 2B (117.2 vs 116.4 MAE) — 4.5× more parameters bought no accuracy on this task,
 suggesting the zero-shot VLM error floor is portion-mass ambiguity, not model capacity.
+MiniMax M3 (a hosted reasoning model, chain-of-thought enabled) reinforces the point from the other
+direction: extensive per-dish reasoning did not help — it lands second-to-last with a strong +104 kcal
+systematic overestimate. Reasoning does not fix calibration.
+
+## Out-of-distribution evaluation (MM-Food-100K, 500 wild food photos)
+
+To measure generalization beyond Nutrition5k plating, both regression models were scored on a seeded
+500-image sample of [MM-Food-100K](https://huggingface.co/datasets/Humanbased-AI/MM-Food-100K)
+(restaurant/home photos; `food_prob ≥ 0.9`). **Caveat: these labels are crowd/AI estimates, not weighed
+measurements** — absolute errors include label noise no model can beat, so the relative comparison is
+the meaningful signal.
+
+| | CalorieK | CalorieCLIP | |
+|---|---|---|---|
+| MAE | **164.9 kcal** | 206.4 kcal | 1.25× better |
+| Median error | **139.1 kcal** | 179.9 kcal | 1.3× better |
+| Within ±100 kcal | **36.0%** | 25.6% | |
+| Bias | **−4.8 kcal** | +98.3 kcal | |
+| R² | **+0.24** | −0.08 | |
+
+Findings, stated honestly:
+- CalorieK beats CalorieCLIP out-of-distribution on every metric, keeps **near-zero bias** on food it
+  has never seen, and retains real signal (R² +0.24) where CalorieCLIP is worse than predicting the
+  mean (R² −0.08).
+- But CalorieK degrades ~3× versus in-distribution (51.4 → 164.9 kcal MAE): it is substantially a
+  specialist in Nutrition5k-style plated dishes. For wild-photo deployment, mixing diverse weakly
+  labeled data into training is the indicated next step.
 
 ## Training method
 
@@ -107,7 +135,7 @@ MODEL_ID = "google/siglip2-base-patch16-224"
 processor = AutoImageProcessor.from_pretrained(MODEL_ID)
 vision = SiglipVisionModel.from_pretrained(MODEL_ID)
 head = NutritionHead()
-state = torch.load(hf_hub_download("Dralkh/CalorieK", "kcalorie.pt"), map_location="cpu", weights_only=False)
+state = torch.load(hf_hub_download("Dralkh/CalorieK", "caloriek.pt"), map_location="cpu", weights_only=False)
 vision.load_state_dict(state["vision_state"]); head.load_state_dict(state["head_state"])
 vision.eval(); head.eval()