organization card: Aspaklarya study overview

README.md

@@ -1,10 +1,75 @@
 ---
-title: README
-emoji: 👁
-colorFrom: yellow
-colorTo: blue
+title: ExperimentField
+emoji: 🔭
+colorFrom: indigo
+colorTo: purple
 sdk: static
 pinned: false
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# Aspaklarya — Qwen3-VL-8B finetuning for remote-sensing VQA + captioning
+
+This organization holds the LoRA adapters and merged checkpoints from a
+systematic finetuning study on **Qwen3-VL-8B-Instruct**, targeting two tasks
+jointly: **visual question answering** and **image captioning** on overhead /
+remote-sensing imagery.
+
+Thirteen experiments were run. Each one changes exactly one thing against a
+fixed reference so the result is attributable.
+
+## Protocol
+
+- **Training**: LoRA via Unsloth, bf16 base, vision tower trainable unless
+  stated otherwise. Each training image contributes one caption example and
+  two VQA examples; caption instructions are sampled from a 30-prompt pool so
+  the model cannot overfit a single phrasing.
+- **Evaluation**: a fixed validation subset, identical for every model and
+  every checkpoint. Each image yields one caption item and one VQA item.
+- **Scoring**: LLM-as-judge grades every answer 0–100 against the reference.
+  Verdicts are cached by item, so re-slicing a result never re-grades it.
+- **Reported slices**: `ALL` (whole eval set) and `DM` (the in-domain subset);
+  caption and VQA are scored separately and the headline is their mean.
+- **Baselines**: stock Qwen3-VL-8B-Instruct, and a previously finetuned
+  in-house model. Every experiment here beats both on every metric.
+
+## Experiments
+
+| # | What was changed | Outcome |
+|---|---|---|
+| **001** | **Reference recipe**: LoRA r=16, α=16, lr 2e-4, linear schedule, vision tuned | **Best overall. Published as the champion.** |
+| 002 | Rank raised to 32 *and* lr halved to 1e-4 | Worse. Confounded two variables; motivated 009 and 013. |
+| 003 | Doubled the training data per image | No gain. Capacity was not the bottleneck. |
+| 004 | Vision tower frozen, language side only | Clearly worse. Visual adaptation matters for this domain. |
+| 005 | Six VQA examples per image + cosine schedule | Best caption of the early runs, but VQA fell. |
+| 006 | Loss reweighting to favour short VQA answers | No gain. Token-count imbalance is not fixable this way. |
+| 007 | Rank lowered to 8 | Slightly worse than r=16. |
+| 008 | Two-stage curriculum: captions first, then VQA | Strongest caption score of the study; VQA never recovered. |
+| 009 | Rank raised to 32, lr held at 2e-4 | Worse. With 002, closes rank as a lever. |
+| 010 | Evaluated on caption prompts never seen in training | Prompt overfitting is mild, but holdout scores reorder runs and must not be compared across pools. |
+| 011 | Regularization: weight decay + gradient clipping + LoRA dropout | Best in-domain score of the study; statistically tied with the champion. |
+| 012 | 011 with LoRA dropout removed | Refuted the idea that dropout was costing VQA. |
+| 013 | Learning rate halved to 1e-4 at the winning rank | Refuted. Lower LR is slower and peaks lower, not higher. |
+
+## What the study concluded
+
+- **Captioning saturates early**; VQA decides every ranking. Caption scores
+  span ~2 points across all thirteen runs, VQA spans ~3.
+- **Model capacity is not the limit.** Rank 8, 16 and 32, doubled data, and
+  resampled VQA all fail to move the ceiling.
+- **The vision tower must be trained** for this imagery.
+- **Measured noise matters.** A paired bootstrap over the eval set puts the
+  run-to-run standard error near 0.6 points, which means the top five
+  experiments are one statistically indistinguishable cluster. Differences
+  below roughly 1.2 points are not real.
+- **Merged models must be re-validated.** Adapter-time and merged-time scores
+  differ slightly, so a checkpoint is only promoted after the merged model is
+  re-scored.
+
+## Repository naming
+
+- `Aspaklarya-<NNN>-step<N>` — LoRA adapter from experiment `NNN`.
+- `Aspaklarya-<NNN>-merged-*` — adapter merged into the bf16 base, ready to
+  serve without PEFT.
+- `Aspaklarya-001-merged-epoch4` — **the published champion.**
+
+All repositories are private.