Blame for README.md

---
title: ExperimentField
emoji: 🔭
colorFrom: indigo
colorTo: purple
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. 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 ... 107 |
| 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 | four-slice seed-1105 draw, 1,991 graded items |
| best | Aspaklarya-020-step4500 | Aspaklarya-101-step4500 |
| what it answered | hyperparameters are saturated, data is the lever | verbose supervision is worth +8 on DM VQA; data is still the only lever |

Both rounds are now scored on the **same four-slice benchmark**, so the
MEAN4 column below is comparable across rounds. The *legacy* and *new VQA*
columns individually are not: 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 |
|---|---|
| Base | Qwen3-VL-8B-Instruct |
| Method | LoRA (Unsloth), bf16 compute; 4-bit base for runs 001-017, **bf16 base from 018 on** |
| Targets | all linear layers, vision + language towers |
| Data | 6,288 images / 18,864 samples for runs 001-019; **11,994 / 35,981 for 020-021** (1 caption + 2 QA each) |
| Caption prompts | sampled from a 30-prompt pool, so the model learns the task not one phrasing |
| Eval | fixed 1000-image draw = 2000 graded items, same seed for every model |
| Scoring | LLM-as-judge, 0-100 per item |
| References | stock Qwen3-VL-8B, plus an in-house finetuned model |

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 |
|---|---|---|---|---|---|---|---|---|
| Aspaklarya-001 | 16 / 16 | 0.0 | 2e-4 | linear | trained | 1c + 2q | 18,864 | reference recipe |
| Aspaklarya-002 | 32 / 64 | 0.05 | 1e-4 | cosine | trained | 1c + 2q | 18,864 | higher rank, lower LR |
| Aspaklarya-003 | 16 / 16 | 0.0 | 2e-4 | linear | trained | 2c + 4q | 37,728 | 2x resampled data |
| Aspaklarya-004 | 16 / 16 | 0.0 | 2e-4 | linear | **frozen** | 1c + 2q | 18,864 | language-only |
| Aspaklarya-005 | 16 / 16 | 0.0 | 2e-4 | cosine | trained | 1c + 6q | 43,864 | VQA-heavy |
| Aspaklarya-006 | 16 / 16 | 0.0 | 2e-4 | linear | trained | 1c + 2q | 18,864 | batch 4 x accum 8 |
| Aspaklarya-007 | 8 / 8 | 0.0 | 2e-4 | linear | trained | 1c + 2q | 18,864 | low rank |
| Aspaklarya-008 | 16 / 16 | 0.0 | 1e-4 | linear | trained | 1c + 4q | 31,440 | more QA, lower LR |
| Aspaklarya-009 | 32 / 32 | 0.0 | 2e-4 | linear | trained | 1c + 2q | 18,864 | high rank |
| Aspaklarya-010 | 16 / 16 | 0.0 | 2e-4 | linear | trained | 1c + 2q | 18,864 | wider prompt pool |
| Aspaklarya-011 | 16 / 16 | 0.05 | 2e-4 | linear | trained | 1c + 2q | 18,864 | dropout, wd 0.05, clip 1.0 |
| Aspaklarya-012 | 16 / 16 | 0.0 | 2e-4 | linear | trained | 1c + 2q | 18,864 | as 011, no dropout |
| Aspaklarya-013 | 16 / 16 | 0.0 | 1e-4 | linear | trained | 1c + 2q | 18,864 | low LR, long schedule |
| Aspaklarya-014 | 16 / 16 | 0.0 | 2e-4 | linear | trained | 1c + 2q | 18,864 | effective batch 16 |
| Aspaklarya-015 | 16 / 16 | 0.0 | 2e-4 | linear | trained | 1c + 2q | 18,864 | effective batch 64 |
| Aspaklarya-016 | 16 / 16 | 0.0 | **4e-4** | linear | trained | 1c + 2q | 18,864 | batch 64 + linear LR scaling |
| Aspaklarya-017 | 16 / 16 | 0.0 | **3e-4** | linear | trained | 1c + 2q | 18,864 | batch 64 + sqrt LR scaling |
| Aspaklarya-018 | 16 / 16 | 0.0 | 2e-4 | linear | trained | 1c + 2q | 18,864 | **bf16 base** (no 4-bit) |
| Aspaklarya-019 | 16 / 16 | 0.05 | 2e-4 | linear | trained | 1c + 2q | 18,864 | **bf16 base** + 011 recipe |
| Aspaklarya-020 | 16 / 16 | 0.0 | 2e-4 | linear | trained | 1c + 2q | **35,981** | bf16 base + **+91% data** |
| Aspaklarya-021 | 16 / 16 | 0.0 | 2e-4 | linear | trained | 1c + 2q | **35,981** | as 020, **8 epochs** not 4 |

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**.

| Model | In-dom. cap | In-dom. vqa | In-dom. overall | All cap | All vqa | All overall | RSICD cap | RSICD vqa | RSICD overall |
|---|---|---|---|---|---|---|---|---|---|
| Aspaklarya-011 | 81.85 ±0.66 | 83.51 ±1.48 | **82.68 ±0.84** | 79.92 ±0.49 | 75.78 ±1.13 | 77.85 ±0.64 | 78.06 ±0.67 | 68.33 ±1.64 | 73.20 ±0.92 |
| Aspaklarya-019 | 82.39 ±0.65 | 82.82 ±1.48 | 82.61 ±0.85 | 80.33 ±0.48 | 75.91 ±1.13 | 78.12 ±0.63 | 78.34 ±0.70 | 69.24 ±1.62 | 73.79 ±0.89 |
| **Aspaklarya-020** | **82.74 ±0.67** | 82.10 ±1.50 | 82.42 ±0.85 | **81.31 ±0.47** | 76.08 ±1.13 | **78.70 ±0.63** | **79.93 ±0.63** | 70.28 ±1.63 | **75.10 ±0.91** |
| Aspaklarya-001 | 81.84 ±0.66 | 83.23 ±1.44 | 82.54 ±0.81 | 79.97 ±0.48 | 75.94 ±1.12 | 77.95 ±0.63 | 78.17 ±0.69 | 68.89 ±1.66 | 73.53 ±0.93 |
| Aspaklarya-018 | 82.28 ±0.69 | 82.70 ±1.51 | 82.49 ±0.86 | 80.46 ±0.49 | 75.80 ±1.11 | 78.13 ±0.63 | 78.70 ±0.67 | 69.15 ±1.61 | 73.92 ±0.90 |
| Aspaklarya-015 | 82.14 ±0.64 | 82.76 ±1.46 | 82.45 ±0.82 | 80.07 ±0.49 | 74.73 ±1.13 | 77.40 ±0.63 | 78.07 ±0.72 | 66.98 ±1.65 | 72.52 ±0.92 |
| Aspaklarya-006 | 81.40 ±0.71 | 82.95 ±1.50 | 82.18 ±0.90 | 79.81 ±0.50 | 75.79 ±1.12 | 77.80 ±0.65 | 78.26 ±0.68 | 68.88 ±1.63 | 73.57 ±0.91 |
| Aspaklarya-021 | 82.14 ±0.70 | 81.93 ±1.53 | 82.03 ±0.88 | 80.93 ±0.48 | 75.41 ±1.13 | 78.17 ±0.65 | 79.76 ±0.64 | 69.12 ±1.65 | 74.44 ±0.92 |
| Aspaklarya-005 | 81.37 ±0.67 | 82.37 ±1.50 | 81.87 ±0.85 | 79.55 ±0.49 | 75.74 ±1.11 | 77.64 ±0.62 | 77.79 ±0.71 | 69.35 ±1.65 | 73.57 ±0.93 |
| Aspaklarya-012 | 81.68 ±0.66 | 82.05 ±1.53 | 81.86 ±0.86 | 79.73 ±0.48 | 75.02 ±1.13 | 77.37 ±0.64 | 77.85 ±0.71 | 68.23 ±1.63 | 73.04 ±0.90 |
| Aspaklarya-014 | 81.51 ±0.67 | 81.93 ±1.48 | 81.72 ±0.83 | 79.95 ±0.49 | 74.77 ±1.13 | 77.36 ±0.63 | 78.45 ±0.71 | 67.87 ±1.66 | 73.16 ±0.92 |
| Aspaklarya-003 | 81.23 ±0.71 | 82.18 ±1.47 | 81.71 ±0.85 | 79.55 ±0.49 | 75.52 ±1.12 | 77.53 ±0.63 | 77.92 ±0.66 | 69.10 ±1.66 | 73.51 ±0.91 |
| Aspaklarya-017 | 81.49 ±0.67 | 81.86 ±1.48 | 81.67 ±0.85 | 79.50 ±0.50 | 75.73 ±1.10 | 77.62 ±0.63 | 77.59 ±0.70 | 69.83 ±1.60 | 73.71 ±0.89 |
| Aspaklarya-001 (ep5) | 80.60 ±0.72 | 82.25 ±1.49 | 81.42 ±0.84 | 79.30 ±0.50 | **76.17 ±1.12** | 77.73 ±0.63 | 78.05 ±0.69 | **70.31 ±1.64** | 74.18 ±0.91 |
| Aspaklarya-013 | 80.58 ±0.69 | 82.20 ±1.48 | 81.39 ±0.84 | 79.11 ±0.49 | 75.35 ±1.10 | 77.23 ±0.61 | 77.70 ±0.69 | 68.74 ±1.60 | 73.22 ±0.90 |
| Aspaklarya-009 | 80.01 ±0.71 | 82.34 ±1.46 | 81.18 ±0.84 | 78.53 ±0.52 | 75.85 ±1.11 | 77.19 ±0.63 | 77.09 ±0.73 | 69.59 ±1.63 | 73.34 ±0.91 |
| Aspaklarya-008 | 80.62 ±0.74 | 81.42 ±1.51 | 81.02 ±0.86 | 79.82 ±0.50 | 75.40 ±1.12 | 77.61 ±0.63 | 79.05 ±0.65 | 69.60 ±1.62 | 74.33 ±0.89 |
| Aspaklarya-007 | 81.13 ±0.71 | 80.82 ±1.51 | 80.97 ±0.86 | 79.33 ±0.51 | 75.33 ±1.12 | 77.33 ±0.64 | 77.60 ±0.70 | 70.03 ±1.62 | 73.81 ±0.90 |
| Aspaklarya-004 | 80.31 ±0.72 | 81.47 ±1.48 | 80.89 ±0.84 | 78.76 ±0.51 | 75.33 ±1.11 | 77.04 ±0.64 | 77.26 ±0.74 | 69.40 ±1.61 | 73.33 ±0.92 |
| Aspaklarya-002 | 80.70 ±0.72 | 80.63 ±1.51 | 80.67 ±0.85 | 79.36 ±0.50 | 74.33 ±1.13 | 76.85 ±0.64 | 78.07 ±0.71 | 68.25 ±1.63 | 73.16 ±0.92 |
| Aspaklarya-010 | 80.50 ±0.74 | 80.66 ±1.58 | 80.58 ±0.90 | 78.88 ±0.52 | 74.28 ±1.13 | 76.58 ±0.64 | 77.32 ±0.72 | 68.12 ±1.65 | 72.72 ±0.91 |
| in-house reference | 71.88 ±1.23 | **83.62 ±1.49** | 77.75 ±1.02 | 65.18 ±0.83 | 73.34 ±1.13 | 69.26 ±0.74 | 58.72 ±1.09 | 63.44 ±1.62 | 61.08 ±1.02 |
| 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.

| LR at batch 64 | ep 1 | ep 2 | ep 3 | ep 4 | ep 5 | Outcome |
|---|---|---|---|---|---|---|
| 2e-4 (015) | 76.92 | 77.03 | 77.25 | 77.64 | **77.92** | peak at epoch 5 |
| 4e-4 (016) | **77.67** | 76.93 | 76.42 | 76.89 | — | stopped at epoch 4, declining |
| 3e-4 (017) | 77.00 | 76.56 | 77.16 | 77.68 | 77.86 | peak at epoch 5, stopped at epoch 6 |

Linear LR scaling is **wrong** 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'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.

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:

| Slice | Caption | VQA | Overall |
|---|---|---|---|
| All (1000 img) | 0.48 | 1.12 | 0.63 |
| In-domain (491) | 0.66 | 1.46 | 0.82 |
| RSICD (509) | 0.69 | 1.64 | 0.91 |

Paired model-vs-model error is smaller because image difficulty cancels: **0.42** overall on the full set, **0.56** in-domain. Use the paired figure for any "A beats B" claim.

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.
- **Captioning is the real gain.** It improves on every slice, by +14 over the base model and +10 to +19 over the in-house reference.
- **The in-domain VQA win is not established.** The in-house reference posts the single highest in-domain VQA score in the table; it collapses outside that domain, where it falls below the untuned base model.
- **The vision tower matters.** Freezing it is the one ablation that consistently hurt.
- **Extra QA per image does not help.** 4-QA and 6-QA mixtures tie with 2-QA.
- **Larger batch trades generality for domain fit.** Effective batch 64 gives the best in-domain captioning in the table and the worst RSICD VQA.
- **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.

Runs **018** and **019** repeat the two strongest recipes with the base loaded in bf16 and nothing else changed. Against the reference, paired on 1000 images:

| run | overall | in-domain | RSICD |
|---|---|---|---|
| Aspaklarya-018 | **+0.17 ±0.42** | -0.05 ±0.59 | **+0.39 ±0.59** |
| Aspaklarya-019 | **+0.16 ±0.40** | **+0.07 ±0.57** | **+0.25 ±0.57** |

These are the first two runs in the sweep to post a positive overall delta, and **018** 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.

The mechanism is not the expected one. Removing quantization error should mostly help the harder task, VQA; instead the gain is **entirely in captioning** (+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.

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.

**Aspaklarya-020** adds them, taking training data to 11,994 images / 35,981 samples (**+91%**). Everything else is identical to 018. Paired against the previous selection on the same 1000 images:

| slice | delta | 95% CI | p(worse) |
|---|---|---|---|
| overall | **+0.74 ±0.47** | [-0.17, 1.66] | 0.059 |
| in-domain | -0.12 ±0.63 | [-1.37, 1.10] | 0.570 |
| **RSICD** | **+1.57 ±0.68** | **[0.22, 2.87]** | **0.012** |
| caption (all) | **+1.34 ±0.42** | — | — |

**This is the first change in the whole programme to clear the error bar.** The RSICD interval excludes zero, and the overall caption gain is 3.2 standard errors. It is also the first result that does **not** trade one slice against the other: in-domain is a clean tie while the general slice moves.

Two further points matter for anyone repeating this:

* **The in-domain penalty is transient.** At matched optimizer steps the extra data looked actively harmful early — step 2,360 was 1.65 behind on in-domain — and that deficit had vanished by step 4,500. Stopping early would have refuted the best run in the programme.
* **The climb to the final checkpoint was schedule, not headroom.** Overall score rose at all seven checkpoints including the last, which suggested an unexhausted budget. **Aspaklarya-021 tested that directly and refuted it** — see below.

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.

| step | overall | caption | vqa | in-domain |
|---|---|---|---|---|
| 1,180 | 77.66 | 79.19 | 76.12 | 81.56 |
| 2,360 | 78.23 | 80.55 | 75.92 | 81.25 |
| 3,540 | 78.37 | 80.33 | 76.41 | 82.18 |
| **4,720** | **78.69** | 80.86 | **76.52** | 82.39 |
| 5,900 | 77.60 | 80.75 | 74.45 | 81.15 |

Its best checkpoint, paired against 020 on the same 1000 images, is negative on every slice and significant on none: overall **-0.53 ±0.43**, in-domain **-0.39 ±0.60**, RSICD **-0.66 ±0.64**.

**Data volume helped; data duration did not.** Four epochs is already the right length at this corpus size.

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 ... 111)

### 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.

### Configurations

All runs share one corpus recipe - `VQA_1126`, 1 caption + 2 QA per image -
and one optimiser: bf16 base, seed 3407, lr 2e-4, linear schedule, warmup
0.03, weight decay 0.001, grad clip 0.3, effective batch 32 (4 x 8), 4 epochs,
max seq 2048, vision tower trained. **One variable each.**

| Model | variable | split | items / images | cost |
|---|---|---|---|---|
| Aspaklarya-101 | **1.9x data** | `train,other` | 35,979 / 11,994 | 9.8 h |
| Aspaklarya-102 | reference | `train` | 18,862 / 6,288 | 5.4 h |
| Aspaklarya-103 | LoRA **rank** 16 -> 32, scale held at 1.0 | `train` | 18,862 / 6,288 | 5.4 h |
| Aspaklarya-104 | LoRA **scale** `alpha/r` 1.0 -> 2.0, rank 16 | `train` | 18,862 / 6,288 | 5.4 h |
| Aspaklarya-105 | **objective balance**, `w_caption` 1.0 -> 0.4 | `train` | 18,862 / 6,288 | 5.4 h |
| Aspaklarya-106 | **LoRA soup** of the four epoch-3 adapters | n/a | n/a | **0 h** |
| Aspaklarya-107 | **cosine** schedule, lr 1e-4, warmup 0.10, 3 epochs | `train` | 18,862 / 6,288 | 3.8 h |
| Aspaklarya-108 | **seed replication** of 102 (1234, 2024) | `train` | 18,862 / 6,288 | 2 x 5.4 h |
| Aspaklarya-109 | **seed replication** of 101 (1234) | `train,other` | 35,979 / 11,994 | 9.9 h |
| Aspaklarya-110 | **seed replication** of 101 (2024) | `train,other` | 35,979 / 11,994 | 11.2 h |
| Aspaklarya-111 | **QA depth** 2 -> 4 per image, caption 1 -> 2 | `train,other` | 71,944 / 11,994 | ~16 h |

### Hyperparameters in full

Every value actually passed to the trainer, read back from each run's
`run_config.json`. Cells that differ from the Aspaklarya-102 reference are
**bold** - that is the single variable the run was built to test.

| | 101 | 102 (ref) | 103 | 104 | 105 | 106 / 106b | 107 | 108 | 109 | 110 | 111 |
|---|---|---|---|---|---|---|---|---|---|---|---|
| split | **`train,other`** | `train` | `train` | `train` | `train` | n/a | `train` | `train` | **`train,other`** | **`train,other`** | **`train,other`** |
| QA per image | 2 | 2 | 2 | 2 | 2 | n/a | 2 | 2 | 2 | 2 | **4** |
| captions per image | 1 | 1 | 1 | 1 | 1 | n/a | 1 | 1 | 1 | 1 | **2** |
| items | **35,979** | 18,862 | 18,862 | 18,862 | 18,862 | n/a | 18,862 | 18,862 | **35,979** | **35,979** | **71,944** |
| images | **11,994** | 6,288 | 6,288 | 6,288 | 6,288 | n/a | 6,288 | 6,288 | **11,994** | **11,994** | **11,994** |
| LoRA rank | 16 | 16 | **32** | 16 | 16 | 64 | 16 | 16 | 16 | 16 | 16 |
| LoRA alpha | 16 | 16 | **32** | **32** | 16 | 64 | 16 | 16 | 16 | 16 | 16 |
| scale `alpha/r` | 1.0 | 1.0 | 1.0 | **2.0** | 1.0 | 1.0 | 1.0 | 1.0 | 1.0 | 1.0 | 1.0 |
| LoRA dropout | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | n/a | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 |
| vision tower | trained | trained | trained | trained | trained | trained | trained | trained | trained | trained | trained |
| `w_caption` | 1.0 | 1.0 | 1.0 | 1.0 | **0.4** | n/a | 1.0 | 1.0 | 1.0 | 1.0 | 1.0 |
| `w_vqa` | 1.0 | 1.0 | 1.0 | 1.0 | 1.0 | n/a | 1.0 | 1.0 | 1.0 | 1.0 | 1.0 |
| loss weighting | token | token | token | token | token | n/a | token | token | token | token | token |
| epochs | 4 | 4 | 4 | 4 | 4 | n/a | **3** | 4 | 4 | 4 | **3** |
| learning rate | 2e-4 | 2e-4 | 2e-4 | 2e-4 | 2e-4 | n/a | **1e-4** | 2e-4 | 2e-4 | 2e-4 | 2e-4 |
| scheduler | linear | linear | linear | linear | linear | n/a | **cosine** | linear | linear | linear | linear |
| warmup ratio | 0.03 | 0.03 | 0.03 | 0.03 | 0.03 | n/a | **0.10** | 0.03 | 0.03 | 0.03 | 0.03 |
| weight decay | 0.001 | 0.001 | 0.001 | 0.001 | 0.001 | n/a | 0.001 | 0.001 | 0.001 | 0.001 | 0.001 |
| grad clip | 0.3 | 0.3 | 0.3 | 0.3 | 0.3 | n/a | 0.3 | 0.3 | 0.3 | 0.3 | 0.3 |
| batch x accum | 4 x 8 | 4 x 8 | 4 x 8 | 4 x 8 | 4 x 8 | n/a | 4 x 8 | 4 x 8 | 4 x 8 | 4 x 8 | 4 x 8 |
| effective batch | 32 | 32 | 32 | 32 | 32 | n/a | 32 | 32 | 32 | 32 | 32 |
| max seq length | 2048 | 2048 | 2048 | 2048 | 2048 | n/a | 2048 | 2048 | 2048 | 2048 | 2048 |
| max image side | 1024 | 1024 | 1024 | 1024 | 1024 | n/a | 1024 | 1024 | 1024 | 1024 | 1024 |
| seed | 3407 | 3407 | 3407 | 3407 | 3407 | n/a | 3407 | **1234 / 2024** | **1234** | **2024** | 3407 |
| base precision | bf16 | bf16 | bf16 | bf16 | bf16 | bf16 | bf16 | bf16 | bf16 | bf16 | bf16 |
| save every | 590 | 295 | 295 | 295 | 295 | n/a | 295 | 295 | 590 | 590 | 1124 |
| steps / epoch | 1125 | 590 | 590 | 590 | 590 | n/a | 590 | 590 | 1125 | 1125 | **2249** |
| total steps | 4500 | 2360 | 2360 | 2360 | 2360 | n/a | 1770 | 2360 | 4500 | 4500 | **6747** |
| wall clock | 9.8 h | 5.4 h | 5.4 h | 5.4 h | 5.4 h | **0 h** | 3.8 h | 2 x 5.4 h | 9.9 h | 11.2 h | ~16 h |

Constant across the whole round and therefore omitted above: base model
`Qwen/Qwen3-VL-8B-Instruct`, no 4-bit quantisation, VQA column `VQA_1126`,
caption prompts drawn from the 50-query bank, eval-loss probe of 400 held-out
items (200 caption / 200 VQA).

**Aspaklarya-106 and 106b are not training runs.** They are exact LoRA soups:
four rank-16 adapters stacked into a single rank-64 adapter, which reproduces
the arithmetic mean of the parent weight deltas to within 2.3e-9. 106 averages
the epoch-3 parents, 106b the epoch-4 parents. Neither consumed GPU training
time; both are listed here so their rank and alpha are not mistaken for a
tuned choice.

**Aspaklarya-108, 109 and 110 are replications, not new configurations.**
108 re-runs 102 at two further seeds; 109 and 110 re-run 101 at one further
seed each. They exist to measure the run-to-run spread that every comparison
on this card silently assumed was small, and to test whether the one positive
result survives a change of seed. They are measurements, not deliverables.

### How these are scored

Round-1XX checkpoints were first ranked on the DM VQA slice alone - the slice
the round was trained for. Every model in the programme has since been
re-judged on **one fixed sample drawn with seed 1105**, drawn once outside the
model loop so all 24 checkpoints see byte-identical items:

| slice | prompt set | new tokens | n |
|---|---|---|---|
| VQA DM | `VQA_1126` (verbose) | 160 | 500 |
| VQA RSICD | legacy terse | 48 | 500 |
| Caption DM | caption-query bank | 320 | 491 |
| Caption RSICD | caption-query bank | 320 | 500 |

`MEAN4` is the unweighted mean, so a model cannot win by specialising on one
slice. Judge: `gemma-4-31b-it-imiai`, greedy, temperature 0, scored 0-100.

Two caveats on the shape of the benchmark. **CAP DM carries 491 items, not
500**, because the underlying 500-image sweep drew 491 DM / 509 RSICD. And
**RSICD VQA has no verbose rewrite** - it is still the legacy question set
generated at a 48-token budget, which clips verbose answers and penalises the
1XX models there.

### Results

**One row per experiment: its best checkpoint.** Every model is scored on the
same 1,991 judged items under the same rubric, so the column is comparable all
the way down. The `ckpts scored` column says how many checkpoints that
experiment contributed; the best of n is an **order statistic**, so it is
biased upward by roughly 0.3-0.5 for n=4 and should not be read as that
recipe's expected quality. Use it to rank recipes, not to size effects.

For that, and for every individual checkpoint, see **Run-to-run sigma** below
and the full 36-row table in `rescore_slices.csv`. Seed replicates of one
recipe (108, 109, 110) are pooled into a single row here.

| Model | VQA DM | VQA RSICD | CAP DM | CAP RSICD | MEAN4 | ckpts scored |
|---|---|---|---|---|---|---|
| **Aspaklarya-101-step4500** | 74.16 | 70.56 | 82.53 | 79.64 | **76.72** | 4 |
| Aspaklarya-109-step3540 | 72.84 | 71.31 | 82.49 | 79.62 | **76.56** | 2 |
| **Aspaklarya-106-soup** | 74.57 | 69.38 | 82.38 | 79.1 | **76.36** | 1 |
| Aspaklarya-110-step3540 | 73.68 | 71.01 | 80.99 | 79.32 | **76.25** | 2 |
| **Aspaklarya-104-step2360** | 73.72 | 69.59 | 82.7 | 78.88 | **76.22** | 4 |
| **Aspaklarya-103-step2360** | 74.13 | 69.82 | 81.34 | 79.3 | **76.15** | 4 |
| Aspaklarya-108-step2360 | 72.66 | 70.83 | 82.1 | 78.58 | **76.04** | 4 |
| **Aspaklarya-105-step1770** | 74.49 | 68.96 | 81.96 | 78.62 | **76.01** | 4 |
| Aspaklarya-106b-soup | 72.77 | 69.99 | 81.53 | 78.73 | **75.75** | 1 |
| **Aspaklarya-102-step1770** | 73.45 | 68.93 | 81.56 | 78.6 | **75.64** | 4 |
| Aspaklarya-107-step1770 | 72.16 | 69.9 | 80.98 | 77.06 | **75.03** | 3 |
| **Aspaklarya-020-step4500** | 66.18 | 70.65 | 82.79 | 79.99 | **74.9** | 1 |
| in-house reference | 65.26 | 63.59 | 71.91 | 58.58 | **64.84** | - |
| Qwen3-VL-8B (base) | 60.64 | 53.13 | 68.31 | 63.28 | **61.34** | - |

### Run-to-run sigma, measured

The previous version of this card said sigma had never been measured and that
no one-point difference should be trusted. That has now been fixed, at the
cost of four replication runs.

**Aspaklarya-108** re-ran the 102 recipe at seeds 1234 and 2024, changing
nothing else. **Aspaklarya-109 and 110** did the same for the 101 recipe, at
seeds 1234 and 2024. With the two original seed-3407 runs that gives two arms
of six scored checkpoints each.

| arm | corpus | checkpoints (MEAN4) | mean | sigma |
|---|---|---|---|---|
| `train,other` (101, 109, 110) | 35,979 | 76.72 · 76.56 · 76.38 · 76.25 · 76.23 · 76.09 | **76.37** | 0.23 |
| `train` (102, 108) | 18,862 | 76.04 · 75.64 · 75.61 · 75.50 · 75.23 · 75.22 | **75.54** | 0.31 |

**sigma = 0.31 MEAN4 for a single checkpoint** (pooled, the wider arm). Two
consequences, and they are not optional:

- A single run targeting less than about **1.0 point** cannot be interpreted.
- The difference between **two** single checkpoints carries SE `0.31 x sqrt(2)
  = 0.44`, so a gap needs roughly **0.9 points** before it means anything, and
  **1.3** before it is comfortable.

Re-reading the whole round against that floor:

| contrast | delta MEAN4 | vs SE 0.44 | verdict |
|---|---|---|---|
| **data scaling** 101/109/110 vs 102/108 | **+0.83** | **5.2 SE** (arms of 6, SE 0.16) | **real** |
| LoRA rank 16 -> 32 (103 vs 102, ep4) | +0.54 | 1.2 SE | not resolved |
| LoRA scale 1.0 -> 2.0 (104 vs 102, ep4) | +0.61 | 1.4 SE | not resolved |
| objective balance `w_caption` 0.4 (105 vs 102, ep4) | -0.02 | 0.0 SE | null |
| cosine + lr 1e-4 (107 vs 102, ep3) | -0.61 | 1.4 SE | not resolved, negative |
| LoRA soup vs mean of its parents (106) | +0.32 | 0.7 SE | not resolved |
| LoRA soup vs mean of its parents (106b) | -0.42 | 1.0 SE | not resolved |

**Only one row of that table survives.** Everything else the round tested is
noise at the resolution the budget bought - which is the honest reading, not a
disappointing one: five levers were eliminated for 25 GPU-hours.

The two soup rows are worth reading together. 106 beat its parents' average
and 106b lost to it, by similar amounts in opposite directions. The defensible
claim is therefore that **souping returns roughly the average of its parents
for zero GPU time**, not that it adds anything.

### What the round actually bought

**Data scaling is the only lever that survived replication.** It is also the
only one that was replicated - and it was replicated specifically because it
was the only candidate worth the GPU time. Three independent seeds of the
large-corpus recipe produced six checkpoints spanning 76.09-76.72; six
small-corpus checkpoints span 75.22-76.04. **The two distributions do not
overlap.** +0.83 at SE 0.16 is 5.2 SE.

Note what this costs in humility about the original claim: 101's headline was
**+1.07**, measured from one seed against one seed. The replicated estimate is
**+0.83**. The effect is real and the first number was 29% too generous - a
clean illustration of why single-seed deltas are quoted here only with a sigma
attached.

**The hyperparameter contrasts are all unresolved or null.** Rank, LoRA scale,
objective balance and schedule shape each moved MEAN4 by less than 1.4 SE.
None of them is worth a further run at this resolution; resolving a 0.5-point
effect would need roughly 6 seeds per arm, or 60 GPU-hours per lever.

**Aspaklarya-020, a full round older, still wins three slices of four.** It
takes VQA RSICD, CAP DM and CAP RSICD, and loses only DM VQA, by 8.0 points.
Round 1XX bought **specialisation onto the rewritten DM question set**, not a
better model overall. Part of that 8.0 is real; part is the 48-token legacy
harness clipping verbose answers.

**The epoch-3 vs epoch-4 question is closed, and the answer is "neither".**
Across the six replicated large-corpus checkpoints the epoch-3/epoch-4 gap is
+0.49, -0.18 and -0.16 depending on seed - it changes sign. Both earlier
claims about a best epoch were reading noise.

**The corpus is still only 27% spent.** `ImageTable.csv` holds 89,119 QA pairs
over 11,994 train+other images, a mean of 7.43 per image, and every run above
used 2. Aspaklarya-111 is the first run to spend more of it, raising QA depth
to 4 while doubling captions to 2 so the 2:1 VQA:caption ratio - and therefore
the objective balance - is held exactly where every previous run had it.

### Aspaklarya-106: an exact LoRA soup

Averaging LoRA adapters naively is wrong, because `mean(B) @ mean(A)` is not
`mean(B @ A)`. The parents are instead **stacked**, so the product reproduces
the mean of the weight deltas exactly:

```
A_cat = [A_1; ...; A_N]            (N*r x k)
B_cat = [c_1*B_1, ..., c_N*B_N]    (d x N*r)
c_i   = (1/N) * (alpha_i/r_i) / (alpha_new/r_new)
```

With `r_new = alpha_new = 64` the reconstruction is exact - verified against
the explicit mean of the parent deltas to **1.9e-9** max absolute error. The
trap is Aspaklarya-104, which carries `alpha/r = 2.0` while its three siblings
carry 1.0: its coefficient must be doubled, and a soup that ignored this would
have silently halved that parent's contribution.

It cost **zero GPU-hours** and lands 2nd of 24.

### What is published here, and what is not

This org holds **22 adapter repos**: Aspaklarya-101 through 105 at four epoch
checkpoints each, Aspaklarya-106-soup, and Aspaklarya-020-step4500 from the
previous round. Five of the runs described above have **no repo at all** -
106b, 107, 108, 109 and 110. That is deliberate, and the reasons differ.

A checkpoint is published when it is **a distinct recipe** and **not strictly
dominated** by something already here. Both tests have to pass.

| run | published | why not |
|---|---|---|
| **106b** (epoch-4 soup) | no | **Strictly dominated.** Same construction as 106-soup, 0.61 lower (75.75 vs 76.36). Shipping a worse version of a model already here is a trap for whoever downloads it, not an extra option. |
| **107** (cosine, lr 1e-4) | no | **Refuted configuration.** 75.03 is the lowest MEAN4 in round 1XX. The run did its job - it ruled the gentler schedule out - and that result is a sentence on this page, not 821 MB of weights. |
| **108** (102 at seeds 1234, 2024) | no | **Calibration, not a deliverable.** |
| **109** (101 at seed 1234) | no | **Calibration, not a deliverable.** |
| **110** (101 at seed 2024) | no | **Calibration, not a deliverable.** |

The last three deserve a fuller answer, because they are the ones that
produced the most useful result on this page.

108, 109 and 110 are not new recipes. **They are the recipes of 102 and 101
re-run with a different random seed and nothing else changed.** Their
adapters are therefore near-duplicates of adapters already published: same
corpus, same rank, same schedule, same everything the config file records.
What they produced is not a model, it is **a number** - sigma 0.31 - and a
number belongs in the text, where it now is.

Publishing them would also do active harm. Six more checkpoints spanning
75.22-76.56 would sit in the listing looking like six more options to choose
between, and the obvious way to choose would be to take the highest one.
That is exactly the mistake the sigma measurement exists to prevent: within
an arm, the spread between seeds **is** the noise, so picking the best seed is
selecting on noise and will not reproduce. The honest artifact is the
published 101 and 102 adapters plus a stated uncertainty on what they score,
not a shelf of lottery tickets.

The cost argument is secondary but real: each adapter is ~821 MB, so the five
unpublished runs would add roughly 5 GB of weights that no one should load in
preference to what is already here.

**Aspaklarya-111 is still training.** It will be published if it clears the
`train,other` arm it is built on - meaning a MEAN4 above roughly 76.8, which
is the arm mean plus two sigma. If it lands inside the arm it will be
reported here as another null and left unpublished, on the same reasoning as
108-110.

### 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.

The legacy slices also reward a terse style the product no longer wants - the
legacy VQA 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.
That is why the legacy numbers are reported but never used to gate a decision.

### Status

- **Aspaklarya-101-step4500 is the round reference** at MEAN4 76.72, but the
  fair description of its recipe is **76.37 +/- 0.23** across three seeds. Any
  of the six large-corpus checkpoints is a defensible choice; picking the
  76.72 one is selecting on noise.
- **Aspaklarya-106-soup is 2nd at 76.36 for no training cost**, and remains
  the better default when GPU time is the constraint - with the caveat above
  that souping buys the parents' average, not a gain over it.
- **Aspaklarya-020-step4500 remains the better general model** (3 slices of 4).
  Choose a 1XX model only when verbose in-domain VQA is what matters.
- **Aspaklarya-106b, 107, 108, 109 and 110 have no adapter repos.** See
  *What is published here, and what is not* above - they are either strictly
  dominated, refuted, or seed replications whose output is a sigma rather
  than a model.
- Open: whether QA depth (111, running) extends the one confirmed lever, and
  whether the `other` split helped by adding **images** or merely **items** -
  111 isolates items, since it adds no new images.