Spaces:

ExperimentField
/

README

Running

App Files Files Community

etzion commited on 4 days ago

Commit

fd297e7

verified ·

1 Parent(s): 0448533

org card: per-experiment hyperparameters and VQA/caption results

Browse files

Files changed (2) hide show

README.md +200 -13
index.html +143 -13

README.md CHANGED Viewed

@@ -7,11 +7,38 @@ sdk: static
 pinned: false
 ---
-# Aspaklarya — Qwen3-VL-8B for remote-sensing VQA + captioning
-LoRA finetunes of **Qwen3-VL-8B-Instruct**, trained jointly on **visual question answering** and **image captioning** over overhead imagery. 21 experiments, 20 of them evaluated (016 was aborted before any checkpoint). All model repos are private.
-## Setup
 | Item | Value |
 |---|---|
@@ -26,7 +53,7 @@ LoRA finetunes of **Qwen3-VL-8B-Instruct**, trained jointly on **visual question
 Winners are merged into the bf16 base and re-validated. For runs 001-017 that merge also removes a train/serve mismatch, since those adapters were fitted on a 4-bit base but are served from bf16 weights; runs 018 onward have no such mismatch.
-## Configurations
 | Model | r / alpha | Dropout | LR | Sched | Vision | Per image | Items | Change |
 |---|---|---|---|---|---|---|---|---|
@@ -54,7 +81,7 @@ Winners are merged into the bf16 base and re-validated. For runs 001-017 that me
 Constant everywhere: effective batch 32 (except 014 at 16, 015-017 at 64), warmup 0.03, grad clip 0.3 (1.0 for 011/012), seed 3407.
-## Results
 Best checkpoint per run on a fixed 1000-image draw (2000 graded items), with bootstrap standard error. **Overall** = mean of caption and VQA. **In-dom.** = the target overhead-imagery subset (491 images); **RSICD** = the public RSICD remote-sensing benchmark subset (509 images). Sorted by in-domain overall; **bold marks the best value in each column**.
@@ -85,7 +112,7 @@ Best checkpoint per run on a fixed 1000-image draw (2000 graded items), with boo
 | Qwen3-VL-8B (base) | 68.37 ±0.89 | 58.94 ±1.50 | 63.66 ±0.92 | 65.68 ±0.66 | 56.01 ±1.01 | 60.84 ±0.62 | 63.07 ±0.94 | 53.18 ±1.46 | 58.13 ±0.91 |
 Every finetune beats both references overall. The selected model, Aspaklarya-020, is **+17.9** over the base model and **+9.4** over the in-house reference on the full set.
-## Learning rate at large batch
 Aspaklarya-015 gave the best captioning in the table at effective batch 64, but it was trained at 2e-4 — the rate tuned for batch 32. Two follow-ups tested whether the batch-64 recipe was simply under-stepped. Scores below come from a separate 500-image draw used for live checkpoint selection, so they compare only with each other.
@@ -99,7 +126,7 @@ Linear LR scaling is **wrong** for LoRA here: 4e-4 peaked at epoch 1 and then lo
 On the shared 1000-image set, neither batch-64 rate beat the reference: 2e-4 is **-0.56** overall and **-0.09** in-domain, 3e-4 is **-0.34** and **-0.87**. Effective batch 64 is therefore refuted at every learning rate tried, and the loss is consistently in VQA while captioning holds or improves. Together with the batch-16 result this closes the batch-size and learning-rate axes: **no hyperparameter lever remains untested**, and further gains have to come from the training data.
-## Error bars
 Bootstrap over images, 3000 resamples. Absolute standard error:
@@ -113,11 +140,11 @@ Paired model-vs-model error is smaller because image difficulty cancels: **0.42*
 Applied exhaustively to the five leading checkpoints (020, 018, 019, 001, 011), the paired bootstrap separates **0 of 10 pairs overall and 0 of 10 in-domain** at 95% — in-domain every pair sits at p ≥ 0.67. The only significant pairs anywhere are three on the RSICD slice, all of them 020 against another model. **The ordering of the results table below the top line is not a measurement.**
-## Caveat
 The hyperparameter runs are statistically tied with each other: among runs 001-019, every confidence interval against the leader straddles zero. **VQA error is ~2.3x caption error**, so no finetune-vs-finetune VQA difference in any column is resolvable; only caption gaps, the gaps to the two references, and the data-scaling result of run 020 are real. Doubling the eval set from 500 to 1000 images did not separate the tied field, and it reshuffled its order.
-## Findings
 - **Hyperparameters are saturated; data is not.** LR, rank, alpha, dropout, weight decay, gradient clipping, batch geometry, effective batch size, LR scaling at large batch and base-weight precision were all swept without beating the noise floor. Adding 91% more training images cleared it on the first attempt.
 - **The slices disagree — but data breaks the trade.** Ranking by in-domain score almost inverts the RSICD ranking, and every hyperparameter lever that moved one slice moved the other the opposite way. Extra data is the exception: it lifted RSICD by 1.57 while leaving in-domain unchanged.
@@ -129,7 +156,7 @@ The hyperparameter runs are statistically tied with each other: among runs 001-0
 - **Training in bf16 instead of 4-bit is a small, consistent positive** — the only change in the whole sweep that never loses on any slice — but it is still inside the error bar.
 - **Data is the remaining lever**, not hyperparameters — and it is not exhausted: the run that used it had not converged when its schedule ended.
-## Base-weight precision
 Runs 001-017 all trained their adapters against **4-bit quantized** base weights, while evaluation and the final merge both use the **bf16** base. That is a train/serve mismatch: the adapter was fitted to compensate for quantization error that is absent at inference time.
@@ -146,7 +173,7 @@ The mechanism is not the expected one. Removing quantization error should mostly
 Removing quantization also costs **~9 GB of GPU memory** during training (19 GB vs 10 GB), which is a real constraint on a shared 40 GB card.
-## Training data
 Every run up to 017 trained on 6,288 images / 18,864 samples. The source table also carries **5,706 further images** that the dataset author left unassigned (`Notes == other`, RSICD only). They are fully populated, statistically indistinguishable from the assigned training rows, and have **zero image-name overlap with either the train or the validation split**, so they are safe to train on.
@@ -168,7 +195,7 @@ Two further points matter for anyone repeating this:
 More data is therefore the live axis, and the only one still producing movement.
-## Schedule length
 **Aspaklarya-021** repeats 020 on the same enlarged corpus with **8 epochs instead of 4** (9,000 steps, checkpoints step-aligned with 020). It peaked at step 4,720 and then declined; the stop rule ended it at step 5,900 with 3,100 steps unspent.
@@ -186,10 +213,170 @@ Its best checkpoint, paired against 020 on the same 1000 images, is negative on
 The failure mode is asymmetric and worth recording: at step 5,900 **captioning was intact (80.75) while VQA collapsed (76.52 → 74.45)**. Overtraining this recipe degrades VQA first, so a stop rule watching captioning alone would not catch it.
-## Status
 The **hyperparameter** search is closed. Rank, alpha, dropout, learning rate, schedule, weight decay, gradient clipping, vision-tower freezing, QA-per-image, caption prompt pool, batch geometry, effective batch size, learning rate at large batch and base-weight precision were all swept across 19 recipes under identical conditions. None displaced the reference recipe by more than the paired error bar.
 Batch size and learning rate illustrate that ceiling. Effective batch 16 came in at **-0.59** overall, batch 64 at 2e-4 at **-0.56**, and batch 64 at 3e-4 at **-0.34** overall and **-0.87** in-domain; batch 64 at 4e-4 was stopped at epoch 4 while declining. In each case captioning held steady or improved while VQA fell.
 Enlarging the training set did what none of those could, and training it longer did not extend the gain. That narrows the remaining question to **data composition**: the corpus added was RSICD-only, and the gain it produced is correspondingly RSICD-only while in-domain stays flat. Additional in-domain imagery is the targeted next step.

 pinned: false
 ---
+# Aspaklarya - Qwen3-VL-8B for remote-sensing VQA + captioning
+LoRA finetunes of **Qwen3-VL-8B-Instruct**, trained jointly on **visual
+question answering** and **image captioning** over overhead imagery. All model
+repos are private.
+The programme has run in **two rounds against two different answer styles**.
+They are not comparable to each other and are not scored on the same
+benchmark.
+| | Round 0XX | Round 1XX |
+|---|---|---|
+| runs | Aspaklarya-001 ... 021 (21, 20 evaluated) | Aspaklarya-101, 102 |
+| VQA column | `VQA` | `VQA_1126` |
+| reference answers | ~4 characters (`yes`, `3`, `river`) | ~300 characters, grounded reasoning |
+| model output | one word | 270-295 characters |
+| benchmark | 1,000-image union draw, 2,000 graded items | 982 questions over 491 in-domain images |
+| best | Aspaklarya-020-step4500 | Aspaklarya-102-step1770 |
+| what it answered | hyperparameters are saturated, data is the lever | verbose supervision is worth +8, extra data is not |
+**Scores from the two rounds must never be ranked against each other.** The
+two benchmarks reward opposite answer styles; see *Why the legacy score
+drops* below.
+## Round 0XX - terse answers (Aspaklarya-001 ... 021)
+The original 21-experiment programme, trained on the one-word `VQA`
+column and scored on the 1,000-image union draw. Everything below this
+heading refers to that round only.
+### Setup
 | Item | Value |
 |---|---|
 Winners are merged into the bf16 base and re-validated. For runs 001-017 that merge also removes a train/serve mismatch, since those adapters were fitted on a 4-bit base but are served from bf16 weights; runs 018 onward have no such mismatch.
+### Configurations
 | Model | r / alpha | Dropout | LR | Sched | Vision | Per image | Items | Change |
 |---|---|---|---|---|---|---|---|---|
 Constant everywhere: effective batch 32 (except 014 at 16, 015-017 at 64), warmup 0.03, grad clip 0.3 (1.0 for 011/012), seed 3407.
+### Results
 Best checkpoint per run on a fixed 1000-image draw (2000 graded items), with bootstrap standard error. **Overall** = mean of caption and VQA. **In-dom.** = the target overhead-imagery subset (491 images); **RSICD** = the public RSICD remote-sensing benchmark subset (509 images). Sorted by in-domain overall; **bold marks the best value in each column**.
 | Qwen3-VL-8B (base) | 68.37 ±0.89 | 58.94 ±1.50 | 63.66 ±0.92 | 65.68 ±0.66 | 56.01 ±1.01 | 60.84 ±0.62 | 63.07 ±0.94 | 53.18 ±1.46 | 58.13 ±0.91 |
 Every finetune beats both references overall. The selected model, Aspaklarya-020, is **+17.9** over the base model and **+9.4** over the in-house reference on the full set.
+### Learning rate at large batch
 Aspaklarya-015 gave the best captioning in the table at effective batch 64, but it was trained at 2e-4 — the rate tuned for batch 32. Two follow-ups tested whether the batch-64 recipe was simply under-stepped. Scores below come from a separate 500-image draw used for live checkpoint selection, so they compare only with each other.
 On the shared 1000-image set, neither batch-64 rate beat the reference: 2e-4 is **-0.56** overall and **-0.09** in-domain, 3e-4 is **-0.34** and **-0.87**. Effective batch 64 is therefore refuted at every learning rate tried, and the loss is consistently in VQA while captioning holds or improves. Together with the batch-16 result this closes the batch-size and learning-rate axes: **no hyperparameter lever remains untested**, and further gains have to come from the training data.
+### Error bars
 Bootstrap over images, 3000 resamples. Absolute standard error:
 Applied exhaustively to the five leading checkpoints (020, 018, 019, 001, 011), the paired bootstrap separates **0 of 10 pairs overall and 0 of 10 in-domain** at 95% — in-domain every pair sits at p ≥ 0.67. The only significant pairs anywhere are three on the RSICD slice, all of them 020 against another model. **The ordering of the results table below the top line is not a measurement.**
+### Caveat
 The hyperparameter runs are statistically tied with each other: among runs 001-019, every confidence interval against the leader straddles zero. **VQA error is ~2.3x caption error**, so no finetune-vs-finetune VQA difference in any column is resolvable; only caption gaps, the gaps to the two references, and the data-scaling result of run 020 are real. Doubling the eval set from 500 to 1000 images did not separate the tied field, and it reshuffled its order.
+### Findings
 - **Hyperparameters are saturated; data is not.** LR, rank, alpha, dropout, weight decay, gradient clipping, batch geometry, effective batch size, LR scaling at large batch and base-weight precision were all swept without beating the noise floor. Adding 91% more training images cleared it on the first attempt.
 - **The slices disagree — but data breaks the trade.** Ranking by in-domain score almost inverts the RSICD ranking, and every hyperparameter lever that moved one slice moved the other the opposite way. Extra data is the exception: it lifted RSICD by 1.57 while leaving in-domain unchanged.
 - **Training in bf16 instead of 4-bit is a small, consistent positive** — the only change in the whole sweep that never loses on any slice — but it is still inside the error bar.
 - **Data is the remaining lever**, not hyperparameters — and it is not exhausted: the run that used it had not converged when its schedule ended.
+### Base-weight precision
 Runs 001-017 all trained their adapters against **4-bit quantized** base weights, while evaluation and the final merge both use the **bf16** base. That is a train/serve mismatch: the adapter was fitted to compensate for quantization error that is absent at inference time.
 Removing quantization also costs **~9 GB of GPU memory** during training (19 GB vs 10 GB), which is a real constraint on a shared 40 GB card.
+### Training data
 Every run up to 017 trained on 6,288 images / 18,864 samples. The source table also carries **5,706 further images** that the dataset author left unassigned (`Notes == other`, RSICD only). They are fully populated, statistically indistinguishable from the assigned training rows, and have **zero image-name overlap with either the train or the validation split**, so they are safe to train on.
 More data is therefore the live axis, and the only one still producing movement.
+### Schedule length
 **Aspaklarya-021** repeats 020 on the same enlarged corpus with **8 epochs instead of 4** (9,000 steps, checkpoints step-aligned with 020). It peaked at step 4,720 and then declined; the stop rule ended it at step 5,900 with 3,100 steps unspent.
 The failure mode is asymmetric and worth recording: at step 5,900 **captioning was intact (80.75) while VQA collapsed (76.52 → 74.45)**. Overtraining this recipe degrades VQA first, so a stop rule watching captioning alone would not catch it.
+### Status
 The **hyperparameter** search is closed. Rank, alpha, dropout, learning rate, schedule, weight decay, gradient clipping, vision-tower freezing, QA-per-image, caption prompt pool, batch geometry, effective batch size, learning rate at large batch and base-weight precision were all swept across 19 recipes under identical conditions. None displaced the reference recipe by more than the paired error bar.
 Batch size and learning rate illustrate that ceiling. Effective batch 16 came in at **-0.59** overall, batch 64 at 2e-4 at **-0.56**, and batch 64 at 3e-4 at **-0.34** overall and **-0.87** in-domain; batch 64 at 4e-4 was stopped at epoch 4 while declining. In each case captioning held steady or improved while VQA fell.
 Enlarging the training set did what none of those could, and training it longer did not extend the gain. That narrows the remaining question to **data composition**: the corpus added was RSICD-only, and the gain it produced is correspondingly RSICD-only while in-domain stays flat. Additional in-domain imagery is the targeted next step.
+## Round 1XX - verbose grounded answers (Aspaklarya-101, 102)
+### Why a second round
+The 0XX round closed with hyperparameters saturated and data identified as
+the only live lever. It also left a product problem: every 0XX model answers
+in one word, because its supervision was one word. `yes` is a correct answer
+to *"are there any anomalies in the scene?"* and a useless one.
+The dataset author then supplied a new column, **`VQA_1126`**, replacing the
+terse answers for the in-domain (DM) images with ~300-character grounded
+answers that name objects, locations and context. The RSICD rows are
+byte-identical to the old column; only the 5,148 DM rows changed, from a mean
+of 5.9 characters to ~304.
+Round 1XX trains on that column and is evaluated on a new benchmark built
+from it: **982 questions over 491 in-domain images**, frozen, same draw for
+every model.
+### Configurations
+Both runs use the Aspaklarya-020 recipe unchanged. The **only** variable
+between them is corpus size.
+| Model | r / alpha | Dropout | LR | Sched | Vision | Per image | Split | Items | Images |
+|---|---|---|---|---|---|---|---|---|---|
+| Aspaklarya-101 | 16 / 16 | 0.0 | 2e-4 | linear | trained | 1c + 2q | `train,other` | 35,979 | 11,994 |
+| Aspaklarya-102 | 16 / 16 | 0.0 | 2e-4 | linear | trained | 1c + 2q | `train` | 18,862 | 6,288 |
+Constant: bf16 base (no 4-bit), effective batch 32 (4 x 8), 4 epochs, warmup
+0.03, weight decay 0.001, grad clip 0.3, max seq 2048, seed 3407. 101 ran
+588 min / 4,500 steps; 102 ran ~320 min / 2,360 steps.
+### Results
+LLM-as-judge (`gemma-4-31b-it-imiai`), greedy, 160 new tokens, 982 items.
+`new` is the VQA_1126 benchmark; `legacy` is the old 2,000-item union draw
+carried over from round 0XX for continuity. **Bold marks the best value.**
+| Model | new | 95% CI | correct (0-5) | concise (0-2) | chars | legacy combined |
+|---|---|---|---|---|---|---|
+| **Aspaklarya-102-step1770** | **74.30** | (71.9, 76.5) | 3.740 | 1.695 | 272 | 74.84 |
+| Aspaklarya-101-step3540 | 74.15 | (71.9, 76.4) | **3.760** | 1.600 | 291 | 75.37 |
+| Aspaklarya-101-step1180 | 74.14 | (71.9, 76.3) | 3.693 | 1.703 | 273 | 74.73 |
+| Aspaklarya-101-step4500 | 74.02 | (71.9, 76.1) | 3.747 | 1.620 | 294 | **75.66** |
+| Aspaklarya-102-step2360 | 73.68 | (71.5, 75.9) | 3.718 | 1.643 | 285 | 74.79 |
+| Aspaklarya-102-step590 | 72.98 | (70.7, 75.2) | 3.675 | 1.663 | 282 | 73.70 |
+| Aspaklarya-101-step2360 | 72.44 | (70.2, 74.6) | 3.655 | 1.634 | 282 | 75.26 |
+| Aspaklarya-102-step1180 | 72.44 | (70.1, 74.7) | 3.637 | 1.675 | 277 | 74.25 |
+| Aspaklarya-011 (0XX) | 66.40 | (64.1, 68.7) | 3.225 | **2.000** | 4.4 | - |
+| Aspaklarya-020 (0XX) | 66.05 | (63.8, 68.2) | 3.200 | **2.000** | 4.1 | 78.83 |
+| in-house reference | 64.78 | (62.5, 67.2) | 3.130 | **2.000** | 4.2 | - |
+| Qwen3-VL-8B (base) | 60.24 | (58.2, 62.3) | 3.346 | 0.634 | 674 | - |
+**Verbose supervision is worth about +8 points** over the best 0XX model on
+the new benchmark, with non-overlapping confidence intervals. It is the
+second change in the whole programme to clear the error bar.
+The stock base model was **re-baselined at 160 new tokens** for this round.
+At the 48-token budget used earlier it scored 53.84, which was a truncation
+artefact; the honest figure is 60.24 and the honest lead over it is +14, not
++20.
+### Data volume does not matter here
+101 and 102 differ only in corpus size: 35,979 items over 11,994 images
+versus 18,862 over 6,288. Best checkpoints land at **74.30 vs 74.15** -
+statistically identical, with 102 nominally ahead on half the data and half
+the GPU time.
+This **does not contradict** the 0XX data-scaling result. The data added in
+both cases is the `other` split, which is RSICD-only, and the round-1XX
+benchmark is in-domain-only. Extra RSICD images lifted the RSICD slice in
+round 0XX and are invisible here. For a DM-targeted product, the larger
+corpus buys nothing.
+### Four epochs are three too many
+Scores by epoch on the new benchmark:
+| run | ep 1 | ep 2 | ep 3 | ep 4 |
+|---|---|---|---|---|
+| Aspaklarya-101 | 74.14 | 72.44 | 74.15 | 74.02 |
+| Aspaklarya-102 | 72.98 | 72.44 | 74.30 | 73.68 |
+Both curves are flat from the first epoch, and every point sits inside every
+other point's interval. The format is learned almost immediately and the
+remaining epochs are noise. **One epoch would have produced the same model**
+for a quarter of the cost - about 7 GPU-hours per run are being spent on
+nothing.
+### Why the legacy score drops
+Every 1XX model scores ~75 on the legacy benchmark against Aspaklarya-020's
+78.83. That is **not a capability regression**. Per-axis, at matched steps:
+| benchmark | model | correct (0-5) | conciseness (0-2) | overall |
+|---|---|---|---|---|
+| legacy (4-char refs) | Aspaklarya-020 | 3.710 | **2.000** | 76.25 |
+| legacy | Aspaklarya-101 | 3.700 | 1.405 | 70.28 |
+| new (300-char refs) | Aspaklarya-020 | 3.200 | **2.000** | 66.05 |
+| new | Aspaklarya-101 | **3.747** | 1.620 | 74.02 |
+On the legacy set the two models are **tied on correctness** (3.700 vs
+3.710) and the entire gap is the conciseness axis. The legacy rubric caps a
+padded-but-correct answer at 89/100 by construction, so a 290-character model
+cannot reach the top band however accurate it is. Captioning, which has long
+references, is unaffected (81.03 vs 81.42).
+On the new set the asymmetry does **not** reverse: Aspaklarya-020 still
+scores a perfect 2.000 conciseness, so it wins the format axis on both
+benchmarks and loses by 8 anyway - on the axis whose rubric explicitly states
+that verbosity is irrelevant.
+### Detail-prompt ablation
+To rule out the alternative explanation - that the 0XX models had the content
+and were merely withholding it - both terse models were re-run on the 982
+questions with an explicit *"answer in detail"* instruction appended.
+| model | prompt | chars | correct (0-5) | new overall |
+|---|---|---|---|---|
+| Aspaklarya-020 | normal | 4.1 | 3.200 | 66.05 |
+| Aspaklarya-020 | detail | 112.3 | 3.249 | 65.31 |
+| in-house reference | normal | 4.2 | 3.130 | 64.78 |
+| in-house reference | detail | 6.9 | 3.026 | 62.97 |
+Aspaklarya-020 complied - 4 to 112 characters - and its correctness did not
+move. The verbosity was elicitable; the content was not there to elicit. The
+in-house reference ignored the instruction almost entirely, which confirms
+the prompt is not doing the work.
+Qualitatively, on *"are there any anomalies in the scene?"*: prompted for
+detail, Aspaklarya-020 produces 450 characters of generic scene description
+that never names an anomaly, while Aspaklarya-101 identifies a specific
+bright irregular object and says why it is out of place.
+**The round-1XX lead is substantive and the legacy dip is cosmetic.**
+### Caveat
+The judge is **text-only**. It compares an answer against a reference and
+never sees the image, so it measures agreement with the reference, not
+verified truth. For a round whose whole premise is longer answers, this
+matters more than it did in round 0XX: the judge can confirm that a model
+supplies reference-matching content, but not that the content is true of the
+pixels. Human spot-checks are the only way to close that gap.
+### Status
+- **Aspaklarya-102-step1770 is the round reference** - equal-best score, half
+  the corpus, half the training time.
+- The legacy benchmark should be treated as **deprecated for this line**. It
+  encodes a terse answer style the product no longer wants, and no single
+  model can lead both benchmarks.
+- Open levers: one-epoch training at a quarter of the cost, in-domain data
+  volume (as opposed to RSICD volume), and a LoRA soup over the eight tied
+  adapters.

index.html CHANGED Viewed

@@ -37,9 +37,26 @@ a { color: #0969da; }
 </head>
 <body>
 <main>
-<h1>Aspaklarya — Qwen3-VL-8B for remote-sensing VQA + captioning</h1>
-<p>LoRA finetunes of <strong>Qwen3-VL-8B-Instruct</strong>, trained jointly on <strong>visual question answering</strong> and <strong>image captioning</strong> over overhead imagery. 21 experiments, 20 of them evaluated (016 was aborted before any checkpoint). All model repos are private.</p>
-<h2>Setup</h2>
 <table style="border-collapse:collapse;margin:1rem 0;font-size:.85rem;border:1px solid #d0d7de"><thead><tr>
 <th style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;background:#f6f8fa;font-weight:600;white-space:nowrap">Item</th>
 <th style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;background:#f6f8fa;font-weight:600;white-space:nowrap">Value</th>
@@ -54,7 +71,7 @@ a { color: #0969da; }
 <tr><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">References</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">stock Qwen3-VL-8B, plus an in-house finetuned model</td></tr>
 </tbody></table>
 <p>Winners are merged into the bf16 base and re-validated. For runs 001-017 that merge also removes a train/serve mismatch, since those adapters were fitted on a 4-bit base but are served from bf16 weights; runs 018 onward have no such mismatch.</p>
-<h2>Configurations</h2>
 <table style="border-collapse:collapse;margin:1rem 0;font-size:.85rem;border:1px solid #d0d7de"><thead><tr>
 <th style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;background:#f6f8fa;font-weight:600;white-space:nowrap">Model</th>
 <th style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;background:#f6f8fa;font-weight:600;white-space:nowrap">r / alpha</th>
@@ -89,7 +106,7 @@ a { color: #0969da; }
 <tr><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">Aspaklarya-021</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">16 / 16</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">0.0</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">2e-4</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">linear</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">trained</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">1c + 2q</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap"><strong>35,981</strong></td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">as 020, <strong>8 epochs</strong> not 4</td></tr>
 </tbody></table>
 <p>Constant everywhere: effective batch 32 (except 014 at 16, 015-017 at 64), warmup 0.03, grad clip 0.3 (1.0 for 011/012), seed 3407.</p>
-<h2>Results</h2>
 <p>Best checkpoint per run on a fixed 1000-image draw (2000 graded items), with bootstrap standard error. <strong>Overall</strong> = mean of caption and VQA. <strong>In-dom.</strong> = the target overhead-imagery subset (491 images); <strong>RSICD</strong> = the public RSICD remote-sensing benchmark subset (509 images). Sorted by in-domain overall; <strong>bold marks the best value in each column</strong>.</p>
 <table style="border-collapse:collapse;margin:1rem 0;font-size:.85rem;border:1px solid #d0d7de"><thead><tr>
 <th style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;background:#f6f8fa;font-weight:600;white-space:nowrap">Model</th>
@@ -128,7 +145,7 @@ a { color: #0969da; }
 <tr><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">Qwen3-VL-8B (base)</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">68.37 ±0.89</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">58.94 ±1.50</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">63.66 ±0.92</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">65.68 ±0.66</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">56.01 ±1.01</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">60.84 ±0.62</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">63.07 ±0.94</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">53.18 ±1.46</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">58.13 ±0.91</td></tr>
 </tbody></table>
 <p>Every finetune beats both references overall. The selected model, Aspaklarya-020, is <strong>+17.9</strong> over the base model and <strong>+9.4</strong> over the in-house reference on the full set.</p>
-<h2>Learning rate at large batch</h2>
 <p>Aspaklarya-015 gave the best captioning in the table at effective batch 64, but it was trained at 2e-4 — the rate tuned for batch 32. Two follow-ups tested whether the batch-64 recipe was simply under-stepped. Scores below come from a separate 500-image draw used for live checkpoint selection, so they compare only with each other.</p>
 <table style="border-collapse:collapse;margin:1rem 0;font-size:.85rem;border:1px solid #d0d7de"><thead><tr>
 <th style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;background:#f6f8fa;font-weight:600;white-space:nowrap">LR at batch 64</th>
@@ -145,7 +162,7 @@ a { color: #0969da; }
 </tbody></table>
 <p>Linear LR scaling is <strong>wrong</strong> for LoRA here: 4e-4 peaked at epoch 1 and then lost ground at every later checkpoint, ending below where 2e-4 started. The adapter&#x27;s effective step size does not follow the full-finetune scaling rule, so raising LR with batch size overshoots rather than compensating. The intermediate rate 3e-4 traced almost exactly the same curve as 2e-4 and peaked at the same epoch, one tenth of a point lower.</p>
 <p>On the shared 1000-image set, neither batch-64 rate beat the reference: 2e-4 is <strong>-0.56</strong> overall and <strong>-0.09</strong> in-domain, 3e-4 is <strong>-0.34</strong> and <strong>-0.87</strong>. Effective batch 64 is therefore refuted at every learning rate tried, and the loss is consistently in VQA while captioning holds or improves. Together with the batch-16 result this closes the batch-size and learning-rate axes: <strong>no hyperparameter lever remains untested</strong>, and further gains have to come from the training data.</p>
-<h2>Error bars</h2>
 <p>Bootstrap over images, 3000 resamples. Absolute standard error:</p>
 <table style="border-collapse:collapse;margin:1rem 0;font-size:.85rem;border:1px solid #d0d7de"><thead><tr>
 <th style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;background:#f6f8fa;font-weight:600;white-space:nowrap">Slice</th>
@@ -159,9 +176,9 @@ a { color: #0969da; }
 </tbody></table>
 <p>Paired model-vs-model error is smaller because image difficulty cancels: <strong>0.42</strong> overall on the full set, <strong>0.56</strong> in-domain. Use the paired figure for any &quot;A beats B&quot; claim.</p>
 <p>Applied exhaustively to the five leading checkpoints (020, 018, 019, 001, 011), the paired bootstrap separates <strong>0 of 10 pairs overall and 0 of 10 in-domain</strong> at 95% — in-domain every pair sits at p ≥ 0.67. The only significant pairs anywhere are three on the RSICD slice, all of them 020 against another model. <strong>The ordering of the results table below the top line is not a measurement.</strong></p>
-<h2>Caveat</h2>
 <p>The hyperparameter runs are statistically tied with each other: among runs 001-019, every confidence interval against the leader straddles zero. <strong>VQA error is ~2.3x caption error</strong>, so no finetune-vs-finetune VQA difference in any column is resolvable; only caption gaps, the gaps to the two references, and the data-scaling result of run 020 are real. Doubling the eval set from 500 to 1000 images did not separate the tied field, and it reshuffled its order.</p>
-<h2>Findings</h2>
 <ul>
 <li><strong>Hyperparameters are saturated; data is not.</strong> LR, rank, alpha, dropout, weight decay, gradient clipping, batch geometry, effective batch size, LR scaling at large batch and base-weight precision were all swept without beating the noise floor. Adding 91% more training images cleared it on the first attempt.</li>
 <li><strong>The slices disagree — but data breaks the trade.</strong> Ranking by in-domain score almost inverts the RSICD ranking, and every hyperparameter lever that moved one slice moved the other the opposite way. Extra data is the exception: it lifted RSICD by 1.57 while leaving in-domain unchanged.</li>
@@ -173,7 +190,7 @@ a { color: #0969da; }
 <li><strong>Training in bf16 instead of 4-bit is a small, consistent positive</strong> — the only change in the whole sweep that never loses on any slice — but it is still inside the error bar.</li>
 <li><strong>Data is the remaining lever</strong>, not hyperparameters — and it is not exhausted: the run that used it had not converged when its schedule ended.</li>
 </ul>
-<h2>Base-weight precision</h2>
 <p>Runs 001-017 all trained their adapters against <strong>4-bit quantized</strong> base weights, while evaluation and the final merge both use the <strong>bf16</strong> base. That is a train/serve mismatch: the adapter was fitted to compensate for quantization error that is absent at inference time.</p>
 <p>Runs <strong>018</strong> and <strong>019</strong> repeat the two strongest recipes with the base loaded in bf16 and nothing else changed. Against the reference, paired on 1000 images:</p>
 <table style="border-collapse:collapse;margin:1rem 0;font-size:.85rem;border:1px solid #d0d7de"><thead><tr>
@@ -188,7 +205,7 @@ a { color: #0969da; }
 <p>These are the first two runs in the sweep to post a positive overall delta, and <strong>018</strong> is the only run that is not worse than the reference on any slice. Both are still well inside the paired error bar (p = 0.34 that each is in fact worse), so neither is promoted.</p>
 <p>The mechanism is not the expected one. Removing quantization error should mostly help the harder task, VQA; instead the gain is <strong>entirely in captioning</strong> (+0.48 and +0.35) while VQA is flat to slightly negative. That is the same caption-up / VQA-flat signature produced by every other lever in the sweep, which suggests precision is acting through the same saturated channel rather than unlocking anything new.</p>
 <p>Removing quantization also costs <strong>~9 GB of GPU memory</strong> during training (19 GB vs 10 GB), which is a real constraint on a shared 40 GB card.</p>
-<h2>Training data</h2>
 <p>Every run up to 017 trained on 6,288 images / 18,864 samples. The source table also carries <strong>5,706 further images</strong> that the dataset author left unassigned (<code>Notes == other</code>, RSICD only). They are fully populated, statistically indistinguishable from the assigned training rows, and have <strong>zero image-name overlap with either the train or the validation split</strong>, so they are safe to train on.</p>
 <p><strong>Aspaklarya-020</strong> adds them, taking training data to 11,994 images / 35,981 samples (<strong>+91%</strong>). Everything else is identical to 018. Paired against the previous selection on the same 1000 images:</p>
 <table style="border-collapse:collapse;margin:1rem 0;font-size:.85rem;border:1px solid #d0d7de"><thead><tr>
@@ -209,7 +226,7 @@ a { color: #0969da; }
 <li><strong>The climb to the final checkpoint was schedule, not headroom.</strong> Overall score rose at all seven checkpoints including the last, which suggested an unexhausted budget. <strong>Aspaklarya-021 tested that directly and refuted it</strong> — see below.</li>
 </ul>
 <p>More data is therefore the live axis, and the only one still producing movement.</p>
-<h2>Schedule length</h2>
 <p><strong>Aspaklarya-021</strong> repeats 020 on the same enlarged corpus with <strong>8 epochs instead of 4</strong> (9,000 steps, checkpoints step-aligned with 020). It peaked at step 4,720 and then declined; the stop rule ended it at step 5,900 with 3,100 steps unspent.</p>
 <table style="border-collapse:collapse;margin:1rem 0;font-size:.85rem;border:1px solid #d0d7de"><thead><tr>
 <th style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;background:#f6f8fa;font-weight:600;white-space:nowrap">step</th>
@@ -227,10 +244,123 @@ a { color: #0969da; }
 <p>Its best checkpoint, paired against 020 on the same 1000 images, is negative on every slice and significant on none: overall <strong>-0.53 ±0.43</strong>, in-domain <strong>-0.39 ±0.60</strong>, RSICD <strong>-0.66 ±0.64</strong>.</p>
 <p><strong>Data volume helped; data duration did not.</strong> Four epochs is already the right length at this corpus size.</p>
 <p>The failure mode is asymmetric and worth recording: at step 5,900 <strong>captioning was intact (80.75) while VQA collapsed (76.52 → 74.45)</strong>. Overtraining this recipe degrades VQA first, so a stop rule watching captioning alone would not catch it.</p>
-<h2>Status</h2>
 <p>The <strong>hyperparameter</strong> search is closed. Rank, alpha, dropout, learning rate, schedule, weight decay, gradient clipping, vision-tower freezing, QA-per-image, caption prompt pool, batch geometry, effective batch size, learning rate at large batch and base-weight precision were all swept across 19 recipes under identical conditions. None displaced the reference recipe by more than the paired error bar.</p>
 <p>Batch size and learning rate illustrate that ceiling. Effective batch 16 came in at <strong>-0.59</strong> overall, batch 64 at 2e-4 at <strong>-0.56</strong>, and batch 64 at 3e-4 at <strong>-0.34</strong> overall and <strong>-0.87</strong> in-domain; batch 64 at 4e-4 was stopped at epoch 4 while declining. In each case captioning held steady or improved while VQA fell.</p>
 <p>Enlarging the training set did what none of those could, and training it longer did not extend the gain. That narrows the remaining question to <strong>data composition</strong>: the corpus added was RSICD-only, and the gain it produced is correspondingly RSICD-only while in-domain stays flat. Additional in-domain imagery is the targeted next step.</p>
 </main>
 </body>
 </html>

 </head>
 <body>
 <main>
+<h1>Aspaklarya - Qwen3-VL-8B for remote-sensing VQA + captioning</h1>
+<p>LoRA finetunes of <strong>Qwen3-VL-8B-Instruct</strong>, trained jointly on <strong>visual question answering</strong> and <strong>image captioning</strong> over overhead imagery. All model repos are private.</p>
+<p>The programme has run in <strong>two rounds against two different answer styles</strong>. They are not comparable to each other and are not scored on the same benchmark.</p>
+<table style="border-collapse:collapse;margin:1rem 0;font-size:.85rem;border:1px solid #d0d7de"><thead><tr>
+<th style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;background:#f6f8fa;font-weight:600;white-space:nowrap"></th>
+<th style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;background:#f6f8fa;font-weight:600;white-space:nowrap">Round 0XX</th>
+<th style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;background:#f6f8fa;font-weight:600;white-space:nowrap">Round 1XX</th>
+</tr></thead><tbody>
+<tr><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">runs</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">Aspaklarya-001 ... 021 (21, 20 evaluated)</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">Aspaklarya-101, 102</td></tr>
+<tr><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">VQA column</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap"><code>VQA</code></td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap"><code>VQA_1126</code></td></tr>
+<tr><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">reference answers</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">~4 characters (<code>yes</code>, <code>3</code>, <code>river</code>)</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">~300 characters, grounded reasoning</td></tr>
+<tr><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">model output</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">one word</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">270-295 characters</td></tr>
+<tr><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">benchmark</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">1,000-image union draw, 2,000 graded items</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">982 questions over 491 in-domain images</td></tr>
+<tr><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">best</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">Aspaklarya-020-step4500</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">Aspaklarya-102-step1770</td></tr>
+<tr><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">what it answered</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">hyperparameters are saturated, data is the lever</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">verbose supervision is worth +8, extra data is not</td></tr>
+</tbody></table>
+<p><strong>Scores from the two rounds must never be ranked against each other.</strong> The two benchmarks reward opposite answer styles; see <em>Why the legacy score drops</em> below.</p>
+<h2>Round 0XX - terse answers (Aspaklarya-001 ... 021)</h2>
+<p>The original 21-experiment programme, trained on the one-word <code>VQA</code> column and scored on the 1,000-image union draw. Everything below this heading refers to that round only.</p>
+<h3>Setup</h3>
 <table style="border-collapse:collapse;margin:1rem 0;font-size:.85rem;border:1px solid #d0d7de"><thead><tr>
 <th style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;background:#f6f8fa;font-weight:600;white-space:nowrap">Item</th>
 <th style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;background:#f6f8fa;font-weight:600;white-space:nowrap">Value</th>
 <tr><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">References</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">stock Qwen3-VL-8B, plus an in-house finetuned model</td></tr>
 </tbody></table>
 <p>Winners are merged into the bf16 base and re-validated. For runs 001-017 that merge also removes a train/serve mismatch, since those adapters were fitted on a 4-bit base but are served from bf16 weights; runs 018 onward have no such mismatch.</p>
+<h3>Configurations</h3>
 <table style="border-collapse:collapse;margin:1rem 0;font-size:.85rem;border:1px solid #d0d7de"><thead><tr>
 <th style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;background:#f6f8fa;font-weight:600;white-space:nowrap">Model</th>
 <th style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;background:#f6f8fa;font-weight:600;white-space:nowrap">r / alpha</th>
 <tr><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">Aspaklarya-021</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">16 / 16</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">0.0</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">2e-4</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">linear</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">trained</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">1c + 2q</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap"><strong>35,981</strong></td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">as 020, <strong>8 epochs</strong> not 4</td></tr>
 </tbody></table>
 <p>Constant everywhere: effective batch 32 (except 014 at 16, 015-017 at 64), warmup 0.03, grad clip 0.3 (1.0 for 011/012), seed 3407.</p>
+<h3>Results</h3>
 <p>Best checkpoint per run on a fixed 1000-image draw (2000 graded items), with bootstrap standard error. <strong>Overall</strong> = mean of caption and VQA. <strong>In-dom.</strong> = the target overhead-imagery subset (491 images); <strong>RSICD</strong> = the public RSICD remote-sensing benchmark subset (509 images). Sorted by in-domain overall; <strong>bold marks the best value in each column</strong>.</p>
 <table style="border-collapse:collapse;margin:1rem 0;font-size:.85rem;border:1px solid #d0d7de"><thead><tr>
 <th style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;background:#f6f8fa;font-weight:600;white-space:nowrap">Model</th>
 <tr><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">Qwen3-VL-8B (base)</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">68.37 ±0.89</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">58.94 ±1.50</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">63.66 ±0.92</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">65.68 ±0.66</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">56.01 ±1.01</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">60.84 ±0.62</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">63.07 ±0.94</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">53.18 ±1.46</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">58.13 ±0.91</td></tr>
 </tbody></table>
 <p>Every finetune beats both references overall. The selected model, Aspaklarya-020, is <strong>+17.9</strong> over the base model and <strong>+9.4</strong> over the in-house reference on the full set.</p>
+<h3>Learning rate at large batch</h3>
 <p>Aspaklarya-015 gave the best captioning in the table at effective batch 64, but it was trained at 2e-4 — the rate tuned for batch 32. Two follow-ups tested whether the batch-64 recipe was simply under-stepped. Scores below come from a separate 500-image draw used for live checkpoint selection, so they compare only with each other.</p>
 <table style="border-collapse:collapse;margin:1rem 0;font-size:.85rem;border:1px solid #d0d7de"><thead><tr>
 <th style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;background:#f6f8fa;font-weight:600;white-space:nowrap">LR at batch 64</th>
 </tbody></table>
 <p>Linear LR scaling is <strong>wrong</strong> for LoRA here: 4e-4 peaked at epoch 1 and then lost ground at every later checkpoint, ending below where 2e-4 started. The adapter&#x27;s effective step size does not follow the full-finetune scaling rule, so raising LR with batch size overshoots rather than compensating. The intermediate rate 3e-4 traced almost exactly the same curve as 2e-4 and peaked at the same epoch, one tenth of a point lower.</p>
 <p>On the shared 1000-image set, neither batch-64 rate beat the reference: 2e-4 is <strong>-0.56</strong> overall and <strong>-0.09</strong> in-domain, 3e-4 is <strong>-0.34</strong> and <strong>-0.87</strong>. Effective batch 64 is therefore refuted at every learning rate tried, and the loss is consistently in VQA while captioning holds or improves. Together with the batch-16 result this closes the batch-size and learning-rate axes: <strong>no hyperparameter lever remains untested</strong>, and further gains have to come from the training data.</p>
+<h3>Error bars</h3>
 <p>Bootstrap over images, 3000 resamples. Absolute standard error:</p>
 <table style="border-collapse:collapse;margin:1rem 0;font-size:.85rem;border:1px solid #d0d7de"><thead><tr>
 <th style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;background:#f6f8fa;font-weight:600;white-space:nowrap">Slice</th>
 </tbody></table>
 <p>Paired model-vs-model error is smaller because image difficulty cancels: <strong>0.42</strong> overall on the full set, <strong>0.56</strong> in-domain. Use the paired figure for any &quot;A beats B&quot; claim.</p>
 <p>Applied exhaustively to the five leading checkpoints (020, 018, 019, 001, 011), the paired bootstrap separates <strong>0 of 10 pairs overall and 0 of 10 in-domain</strong> at 95% — in-domain every pair sits at p ≥ 0.67. The only significant pairs anywhere are three on the RSICD slice, all of them 020 against another model. <strong>The ordering of the results table below the top line is not a measurement.</strong></p>
+<h3>Caveat</h3>
 <p>The hyperparameter runs are statistically tied with each other: among runs 001-019, every confidence interval against the leader straddles zero. <strong>VQA error is ~2.3x caption error</strong>, so no finetune-vs-finetune VQA difference in any column is resolvable; only caption gaps, the gaps to the two references, and the data-scaling result of run 020 are real. Doubling the eval set from 500 to 1000 images did not separate the tied field, and it reshuffled its order.</p>
+<h3>Findings</h3>
 <ul>
 <li><strong>Hyperparameters are saturated; data is not.</strong> LR, rank, alpha, dropout, weight decay, gradient clipping, batch geometry, effective batch size, LR scaling at large batch and base-weight precision were all swept without beating the noise floor. Adding 91% more training images cleared it on the first attempt.</li>
 <li><strong>The slices disagree — but data breaks the trade.</strong> Ranking by in-domain score almost inverts the RSICD ranking, and every hyperparameter lever that moved one slice moved the other the opposite way. Extra data is the exception: it lifted RSICD by 1.57 while leaving in-domain unchanged.</li>
 <li><strong>Training in bf16 instead of 4-bit is a small, consistent positive</strong> — the only change in the whole sweep that never loses on any slice — but it is still inside the error bar.</li>
 <li><strong>Data is the remaining lever</strong>, not hyperparameters — and it is not exhausted: the run that used it had not converged when its schedule ended.</li>
 </ul>
+<h3>Base-weight precision</h3>
 <p>Runs 001-017 all trained their adapters against <strong>4-bit quantized</strong> base weights, while evaluation and the final merge both use the <strong>bf16</strong> base. That is a train/serve mismatch: the adapter was fitted to compensate for quantization error that is absent at inference time.</p>
 <p>Runs <strong>018</strong> and <strong>019</strong> repeat the two strongest recipes with the base loaded in bf16 and nothing else changed. Against the reference, paired on 1000 images:</p>
 <table style="border-collapse:collapse;margin:1rem 0;font-size:.85rem;border:1px solid #d0d7de"><thead><tr>
 <p>These are the first two runs in the sweep to post a positive overall delta, and <strong>018</strong> is the only run that is not worse than the reference on any slice. Both are still well inside the paired error bar (p = 0.34 that each is in fact worse), so neither is promoted.</p>
 <p>The mechanism is not the expected one. Removing quantization error should mostly help the harder task, VQA; instead the gain is <strong>entirely in captioning</strong> (+0.48 and +0.35) while VQA is flat to slightly negative. That is the same caption-up / VQA-flat signature produced by every other lever in the sweep, which suggests precision is acting through the same saturated channel rather than unlocking anything new.</p>
 <p>Removing quantization also costs <strong>~9 GB of GPU memory</strong> during training (19 GB vs 10 GB), which is a real constraint on a shared 40 GB card.</p>
+<h3>Training data</h3>
 <p>Every run up to 017 trained on 6,288 images / 18,864 samples. The source table also carries <strong>5,706 further images</strong> that the dataset author left unassigned (<code>Notes == other</code>, RSICD only). They are fully populated, statistically indistinguishable from the assigned training rows, and have <strong>zero image-name overlap with either the train or the validation split</strong>, so they are safe to train on.</p>
 <p><strong>Aspaklarya-020</strong> adds them, taking training data to 11,994 images / 35,981 samples (<strong>+91%</strong>). Everything else is identical to 018. Paired against the previous selection on the same 1000 images:</p>
 <table style="border-collapse:collapse;margin:1rem 0;font-size:.85rem;border:1px solid #d0d7de"><thead><tr>
 <li><strong>The climb to the final checkpoint was schedule, not headroom.</strong> Overall score rose at all seven checkpoints including the last, which suggested an unexhausted budget. <strong>Aspaklarya-021 tested that directly and refuted it</strong> — see below.</li>
 </ul>
 <p>More data is therefore the live axis, and the only one still producing movement.</p>
+<h3>Schedule length</h3>
 <p><strong>Aspaklarya-021</strong> repeats 020 on the same enlarged corpus with <strong>8 epochs instead of 4</strong> (9,000 steps, checkpoints step-aligned with 020). It peaked at step 4,720 and then declined; the stop rule ended it at step 5,900 with 3,100 steps unspent.</p>
 <table style="border-collapse:collapse;margin:1rem 0;font-size:.85rem;border:1px solid #d0d7de"><thead><tr>
 <th style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;background:#f6f8fa;font-weight:600;white-space:nowrap">step</th>
 <p>Its best checkpoint, paired against 020 on the same 1000 images, is negative on every slice and significant on none: overall <strong>-0.53 ±0.43</strong>, in-domain <strong>-0.39 ±0.60</strong>, RSICD <strong>-0.66 ±0.64</strong>.</p>
 <p><strong>Data volume helped; data duration did not.</strong> Four epochs is already the right length at this corpus size.</p>
 <p>The failure mode is asymmetric and worth recording: at step 5,900 <strong>captioning was intact (80.75) while VQA collapsed (76.52 → 74.45)</strong>. Overtraining this recipe degrades VQA first, so a stop rule watching captioning alone would not catch it.</p>
+<h3>Status</h3>
 <p>The <strong>hyperparameter</strong> search is closed. Rank, alpha, dropout, learning rate, schedule, weight decay, gradient clipping, vision-tower freezing, QA-per-image, caption prompt pool, batch geometry, effective batch size, learning rate at large batch and base-weight precision were all swept across 19 recipes under identical conditions. None displaced the reference recipe by more than the paired error bar.</p>
 <p>Batch size and learning rate illustrate that ceiling. Effective batch 16 came in at <strong>-0.59</strong> overall, batch 64 at 2e-4 at <strong>-0.56</strong>, and batch 64 at 3e-4 at <strong>-0.34</strong> overall and <strong>-0.87</strong> in-domain; batch 64 at 4e-4 was stopped at epoch 4 while declining. In each case captioning held steady or improved while VQA fell.</p>
 <p>Enlarging the training set did what none of those could, and training it longer did not extend the gain. That narrows the remaining question to <strong>data composition</strong>: the corpus added was RSICD-only, and the gain it produced is correspondingly RSICD-only while in-domain stays flat. Additional in-domain imagery is the targeted next step.</p>
+<h2>Round 1XX - verbose grounded answers (Aspaklarya-101, 102)</h2>
+<h3>Why a second round</h3>
+<p>The 0XX round closed with hyperparameters saturated and data identified as the only live lever. It also left a product problem: every 0XX model answers in one word, because its supervision was one word. <code>yes</code> is a correct answer to <em>&quot;are there any anomalies in the scene?&quot;</em> and a useless one.</p>
+<p>The dataset author then supplied a new column, <strong><code>VQA_1126</code></strong>, replacing the terse answers for the in-domain (DM) images with ~300-character grounded answers that name objects, locations and context. The RSICD rows are byte-identical to the old column; only the 5,148 DM rows changed, from a mean of 5.9 characters to ~304.</p>
+<p>Round 1XX trains on that column and is evaluated on a new benchmark built from it: <strong>982 questions over 491 in-domain images</strong>, frozen, same draw for every model.</p>
+<h3>Configurations</h3>
+<p>Both runs use the Aspaklarya-020 recipe unchanged. The <strong>only</strong> variable between them is corpus size.</p>
+<table style="border-collapse:collapse;margin:1rem 0;font-size:.85rem;border:1px solid #d0d7de"><thead><tr>
+<th style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;background:#f6f8fa;font-weight:600;white-space:nowrap">Model</th>
+<th style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;background:#f6f8fa;font-weight:600;white-space:nowrap">r / alpha</th>
+<th style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;background:#f6f8fa;font-weight:600;white-space:nowrap">Dropout</th>
+<th style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;background:#f6f8fa;font-weight:600;white-space:nowrap">LR</th>
+<th style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;background:#f6f8fa;font-weight:600;white-space:nowrap">Sched</th>
+<th style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;background:#f6f8fa;font-weight:600;white-space:nowrap">Vision</th>
+<th style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;background:#f6f8fa;font-weight:600;white-space:nowrap">Per image</th>
+<th style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;background:#f6f8fa;font-weight:600;white-space:nowrap">Split</th>
+<th style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;background:#f6f8fa;font-weight:600;white-space:nowrap">Items</th>
+<th style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;background:#f6f8fa;font-weight:600;white-space:nowrap">Images</th>
+</tr></thead><tbody>
+<tr><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">Aspaklarya-101</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">16 / 16</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">0.0</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">2e-4</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">linear</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">trained</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">1c + 2q</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap"><code>train,other</code></td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">35,979</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">11,994</td></tr>
+<tr><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">Aspaklarya-102</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">16 / 16</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">0.0</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">2e-4</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">linear</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">trained</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">1c + 2q</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap"><code>train</code></td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">18,862</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">6,288</td></tr>
+</tbody></table>
+<p>Constant: bf16 base (no 4-bit), effective batch 32 (4 x 8), 4 epochs, warmup 0.03, weight decay 0.001, grad clip 0.3, max seq 2048, seed 3407. 101 ran 588 min / 4,500 steps; 102 ran ~320 min / 2,360 steps.</p>
+<h3>Results</h3>
+<p>LLM-as-judge (<code>gemma-4-31b-it-imiai</code>), greedy, 160 new tokens, 982 items. <code>new</code> is the VQA_1126 benchmark; <code>legacy</code> is the old 2,000-item union draw carried over from round 0XX for continuity. <strong>Bold marks the best value.</strong></p>
+<table style="border-collapse:collapse;margin:1rem 0;font-size:.85rem;border:1px solid #d0d7de"><thead><tr>
+<th style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;background:#f6f8fa;font-weight:600;white-space:nowrap">Model</th>
+<th style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;background:#f6f8fa;font-weight:600;white-space:nowrap">new</th>
+<th style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;background:#f6f8fa;font-weight:600;white-space:nowrap">95% CI</th>
+<th style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;background:#f6f8fa;font-weight:600;white-space:nowrap">correct (0-5)</th>
+<th style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;background:#f6f8fa;font-weight:600;white-space:nowrap">concise (0-2)</th>
+<th style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;background:#f6f8fa;font-weight:600;white-space:nowrap">chars</th>
+<th style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;background:#f6f8fa;font-weight:600;white-space:nowrap">legacy combined</th>
+</tr></thead><tbody>
+<tr><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap"><strong>Aspaklarya-102-step1770</strong></td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap"><strong>74.30</strong></td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">(71.9, 76.5)</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">3.740</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">1.695</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">272</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">74.84</td></tr>
+<tr><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">Aspaklarya-101-step3540</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">74.15</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">(71.9, 76.4)</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap"><strong>3.760</strong></td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">1.600</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">291</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">75.37</td></tr>
+<tr><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">Aspaklarya-101-step1180</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">74.14</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">(71.9, 76.3)</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">3.693</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">1.703</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">273</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">74.73</td></tr>
+<tr><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">Aspaklarya-101-step4500</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">74.02</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">(71.9, 76.1)</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">3.747</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">1.620</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">294</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap"><strong>75.66</strong></td></tr>
+<tr><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">Aspaklarya-102-step2360</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">73.68</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">(71.5, 75.9)</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">3.718</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">1.643</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">285</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">74.79</td></tr>
+<tr><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">Aspaklarya-102-step590</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">72.98</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">(70.7, 75.2)</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">3.675</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">1.663</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">282</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">73.70</td></tr>
+<tr><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">Aspaklarya-101-step2360</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">72.44</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">(70.2, 74.6)</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">3.655</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">1.634</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">282</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">75.26</td></tr>
+<tr><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">Aspaklarya-102-step1180</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">72.44</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">(70.1, 74.7)</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">3.637</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">1.675</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">277</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">74.25</td></tr>
+<tr><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">Aspaklarya-011 (0XX)</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">66.40</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">(64.1, 68.7)</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">3.225</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap"><strong>2.000</strong></td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">4.4</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">-</td></tr>
+<tr><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">Aspaklarya-020 (0XX)</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">66.05</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">(63.8, 68.2)</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">3.200</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap"><strong>2.000</strong></td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">4.1</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">78.83</td></tr>
+<tr><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">in-house reference</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">64.78</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">(62.5, 67.2)</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">3.130</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap"><strong>2.000</strong></td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">4.2</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">-</td></tr>
+<tr><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">Qwen3-VL-8B (base)</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">60.24</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">(58.2, 62.3)</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">3.346</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">0.634</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">674</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">-</td></tr>
+</tbody></table>
+<p><strong>Verbose supervision is worth about +8 points</strong> over the best 0XX model on the new benchmark, with non-overlapping confidence intervals. It is the second change in the whole programme to clear the error bar.</p>
+<p>The stock base model was <strong>re-baselined at 160 new tokens</strong> for this round. At the 48-token budget used earlier it scored 53.84, which was a truncation artefact; the honest figure is 60.24 and the honest lead over it is +14, not +20.</p>
+<h3>Data volume does not matter here</h3>
+<p>101 and 102 differ only in corpus size: 35,979 items over 11,994 images versus 18,862 over 6,288. Best checkpoints land at <strong>74.30 vs 74.15</strong> - statistically identical, with 102 nominally ahead on half the data and half the GPU time.</p>
+<p>This <strong>does not contradict</strong> the 0XX data-scaling result. The data added in both cases is the <code>other</code> split, which is RSICD-only, and the round-1XX benchmark is in-domain-only. Extra RSICD images lifted the RSICD slice in round 0XX and are invisible here. For a DM-targeted product, the larger corpus buys nothing.</p>
+<h3>Four epochs are three too many</h3>
+<p>Scores by epoch on the new benchmark:</p>
+<table style="border-collapse:collapse;margin:1rem 0;font-size:.85rem;border:1px solid #d0d7de"><thead><tr>
+<th style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;background:#f6f8fa;font-weight:600;white-space:nowrap">run</th>
+<th style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;background:#f6f8fa;font-weight:600;white-space:nowrap">ep 1</th>
+<th style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;background:#f6f8fa;font-weight:600;white-space:nowrap">ep 2</th>
+<th style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;background:#f6f8fa;font-weight:600;white-space:nowrap">ep 3</th>
+<th style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;background:#f6f8fa;font-weight:600;white-space:nowrap">ep 4</th>
+</tr></thead><tbody>
+<tr><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">Aspaklarya-101</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">74.14</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">72.44</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">74.15</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">74.02</td></tr>
+<tr><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">Aspaklarya-102</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">72.98</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">72.44</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">74.30</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">73.68</td></tr>
+</tbody></table>
+<p>Both curves are flat from the first epoch, and every point sits inside every other point&#x27;s interval. The format is learned almost immediately and the remaining epochs are noise. <strong>One epoch would have produced the same model</strong> for a quarter of the cost - about 7 GPU-hours per run are being spent on nothing.</p>
+<h3>Why the legacy score drops</h3>
+<p>Every 1XX model scores ~75 on the legacy benchmark against Aspaklarya-020&#x27;s 78.83. That is <strong>not a capability regression</strong>. Per-axis, at matched steps:</p>
+<table style="border-collapse:collapse;margin:1rem 0;font-size:.85rem;border:1px solid #d0d7de"><thead><tr>
+<th style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;background:#f6f8fa;font-weight:600;white-space:nowrap">benchmark</th>
+<th style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;background:#f6f8fa;font-weight:600;white-space:nowrap">model</th>
+<th style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;background:#f6f8fa;font-weight:600;white-space:nowrap">correct (0-5)</th>
+<th style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;background:#f6f8fa;font-weight:600;white-space:nowrap">conciseness (0-2)</th>
+<th style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;background:#f6f8fa;font-weight:600;white-space:nowrap">overall</th>
+</tr></thead><tbody>
+<tr><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">legacy (4-char refs)</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">Aspaklarya-020</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">3.710</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap"><strong>2.000</strong></td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">76.25</td></tr>
+<tr><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">legacy</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">Aspaklarya-101</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">3.700</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">1.405</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">70.28</td></tr>
+<tr><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">new (300-char refs)</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">Aspaklarya-020</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">3.200</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap"><strong>2.000</strong></td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">66.05</td></tr>
+<tr><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">new</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">Aspaklarya-101</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap"><strong>3.747</strong></td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">1.620</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">74.02</td></tr>
+</tbody></table>
+<p>On the legacy set the two models are <strong>tied on correctness</strong> (3.700 vs 3.710) and the entire gap is the conciseness axis. The legacy rubric caps a padded-but-correct answer at 89/100 by construction, so a 290-character model cannot reach the top band however accurate it is. Captioning, which has long references, is unaffected (81.03 vs 81.42).</p>
+<p>On the new set the asymmetry does <strong>not</strong> reverse: Aspaklarya-020 still scores a perfect 2.000 conciseness, so it wins the format axis on both benchmarks and loses by 8 anyway - on the axis whose rubric explicitly states that verbosity is irrelevant.</p>
+<h3>Detail-prompt ablation</h3>
+<p>To rule out the alternative explanation - that the 0XX models had the content and were merely withholding it - both terse models were re-run on the 982 questions with an explicit <em>&quot;answer in detail&quot;</em> instruction appended.</p>
+<table style="border-collapse:collapse;margin:1rem 0;font-size:.85rem;border:1px solid #d0d7de"><thead><tr>
+<th style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;background:#f6f8fa;font-weight:600;white-space:nowrap">model</th>
+<th style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;background:#f6f8fa;font-weight:600;white-space:nowrap">prompt</th>
+<th style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;background:#f6f8fa;font-weight:600;white-space:nowrap">chars</th>
+<th style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;background:#f6f8fa;font-weight:600;white-space:nowrap">correct (0-5)</th>
+<th style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;background:#f6f8fa;font-weight:600;white-space:nowrap">new overall</th>
+</tr></thead><tbody>
+<tr><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">Aspaklarya-020</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">normal</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">4.1</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">3.200</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">66.05</td></tr>
+<tr><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">Aspaklarya-020</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">detail</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">112.3</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">3.249</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">65.31</td></tr>
+<tr><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">in-house reference</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">normal</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">4.2</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">3.130</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">64.78</td></tr>
+<tr><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">in-house reference</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">detail</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">6.9</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">3.026</td><td style="border:1px solid #d0d7de;padding:5px 10px;text-align:left;white-space:nowrap">62.97</td></tr>
+</tbody></table>
+<p>Aspaklarya-020 complied - 4 to 112 characters - and its correctness did not move. The verbosity was elicitable; the content was not there to elicit. The in-house reference ignored the instruction almost entirely, which confirms the prompt is not doing the work.</p>
+<p>Qualitatively, on <em>&quot;are there any anomalies in the scene?&quot;</em>: prompted for detail, Aspaklarya-020 produces 450 characters of generic scene description that never names an anomaly, while Aspaklarya-101 identifies a specific bright irregular object and says why it is out of place.</p>
+<p><strong>The round-1XX lead is substantive and the legacy dip is cosmetic.</strong></p>
+<h3>Caveat</h3>
+<p>The judge is <strong>text-only</strong>. It compares an answer against a reference and never sees the image, so it measures agreement with the reference, not verified truth. For a round whose whole premise is longer answers, this matters more than it did in round 0XX: the judge can confirm that a model supplies reference-matching content, but not that the content is true of the pixels. Human spot-checks are the only way to close that gap.</p>
+<h3>Status</h3>
+<ul>
+<li><strong>Aspaklarya-102-step1770 is the round reference</strong> - equal-best score, half</li>
+</ul>
+<p>  the corpus, half the training time.</p>
+<ul>
+<li>The legacy benchmark should be treated as <strong>deprecated for this line</strong>. It</li>
+</ul>
+<p>  encodes a terse answer style the product no longer wants, and no single   model can lead both benchmarks.</p>
+<ul>
+<li>Open levers: one-epoch training at a quarter of the cost, in-domain data</li>
+</ul>
+<p>  volume (as opposed to RSICD volume), and a LoRA soup over the eight tied   adapters.</p>
 </main>
 </body>
 </html>