Mac LLM/VLM Post-training Testing · visual question answering (SpatialEval spatial-map)
November 19, 2026 · Qwen3.5-0.8B and 2B with LoRA · Apple M5 MacBook Pro, 16 GB · 9 runs, 13 hours of training and evaluation
Code on GitHub · GSM8K text post-training · math-story post-training · Inference benchmark
The text-only studies showed what SFT and RL do to a small model on a MacBook. Does the same recipe work when the model has to look at an image, and what does it cost the machine?
We used MilaWang/SpatialEval, spatial-map split: each item is a picture of a map with named locations and a multiple-choice question about it, in three types: direction ("in which direction is X relative to Y?"), which object ("which object is in the northwest of Y?") and count ("how many objects are to the east of Y?"). We trained on a small slice, 500 items, and tested on 100 held-out items, because the point is to benchmark what the Mac does, not to chase a leaderboard. The same six recipes as before ran on Qwen3.5-0.8B; three of them were repeated on Qwen3.5-2B. All use LoRA (rank 16) on the language model, with the vision tower frozen.
Images go through a frozen vision tower once; everything we train is the language model.
The answer is: C. Northwest, lr 2e-5, 5 epochs), gentle SFT (lr 5e-6, 2 epochs), self-distilled SFT (rejection-sample the model's own correct answers, then train on them), and three critic-free RL methods: GRPO, RLOO and GSPO (4 prompts × 4 sampled answers per step, temperature 1.0, correctness-only reward that also needs the answer line, no KL penalty, lr 2e-6, 5 epochs = 625 steps).The answer is: X line with the right letter. This is strict on purpose: it is also the RL reward.The 0.8B model scores 39% and its biggest problem is counting.
| What the untrained model did on the 100 test questions | Count |
|---|---|
| correct | 39 |
| wrong letter | 30 |
| cut off at the token cap | 29 |
| finished, no 'The answer is: X' line | 2 |
Direction and which-object questions are only partly answered correctly (61% and 57%); count questions are almost never answered correctly (8%), and 41% of them ran into the token cap while the model enumerated objects one by one. Chance is 25% on four options, and the most common gold letter covers 28% of the test set, so a model that only guessed the most common letter would score about 28%.
Five epochs of answer-only SFT lifts the 0.8B model from 39% to 89%.
The SFT targets contain only the final line, taken from the dataset's oracle option. That is a deliberate simplification: there are no reference reasoning traces for these maps, so SFT teaches the model to answer directly. It works because the model already has the visual skill; it only needed to commit.
The aggressive recipe (lr 2e-5, 5 epochs) climbs slowly (39%, 47%, 64%, 76%, 86%, 89%); the gentle recipe (lr 5e-6, 2 epochs) gets to 91% in well under half the time. Self-distilled SFT ends at only 63%: it trains on the model's own correct answers, so it can only reinforce what the model already did, including the rambling style, and its training set was thin on the hard type: only 19% of its examples were count questions (88 of 457), against 39% of the test set, because the untrained model rarely got those right.
GRPO, RLOO and GSPO need no reference answers, only the correct-or-not reward.
After the first epoch the three RL runs scored GRPO 78%, RLOO 75%, GSPO 84%, and finished at GRPO 84%, RLOO 88%, GSPO 87%. As in the text studies, RL also changes how the model writes, with no length penalty: shorter answers finish more often and so earn more reward.
By the end, 81% of GRPO's groups, 85% of RLOO's groups, 96% of GSPO's groups carried no signal, which suggests the later epochs did little (we did not test shorter training).
| Run (0.8B, LoRA) | Training | Direction | Which object | Count | All | Tokens / answer | Truncated | Hours |
|---|---|---|---|---|---|---|---|---|
| Untrained | – | 61% | 57% | 8% | 39% | 392 | 29% | – |
| SFT | 5 ep | 94% | 90% | 85% | 89% | 8 | 0% | 0.5 |
| SFT gentle | 2 ep | 94% | 97% | 85% | 91% | 8 | 0% | 0.2 |
| SFT self-distilled | 2 ep | 77% | 70% | 46% | 63% | 302 | 12% | 0.4 |
| GRPO | 5 ep | 74% | 90% | 87% | 84% | 158 | 0% | 3.7 |
| RLOO | 5 ep | 90% | 87% | 87% | 88% | 87 | 0% | 2.6 |
| GSPO | 5 ep | 94% | 90% | 79% | 87% | 38 | 1% | 1.5 |
Slightly better untrained, equal after training.
| Run (2B, LoRA) | Training | Direction | Which object | Count | All | Tokens / answer | Truncated | Hours |
|---|---|---|---|---|---|---|---|---|
| Untrained | – | 74% | 77% | 8% | 49% | 322 | 14% | – |
| SFT | 5 ep | 100% | 90% | 82% | 90% | 8 | 0% | 0.7 |
| SFT gentle | 2 ep | 97% | 93% | 82% | 90% | 8 | 0% | 0.3 |
| GRPO | 5 ep | 90% | 83% | 82% | 85% | 53 | 0% | 3.2 |
| Method | 0.8B: untrained → final | 2B: untrained → final | 0.8B hours | 2B hours |
|---|---|---|---|---|
| SFT | 39% → 89% | 49% → 90% | 0.5 | 0.7 |
| SFT gentle | 39% → 91% | 49% → 90% | 0.2 | 0.3 |
| GRPO | 39% → 84% | 49% → 85% | 3.7 | 3.2 |
The 2B model starts higher (49% vs 39%), but after SFT both reach about 89%-90%, and 2B GRPO ended at 85% (best epoch 89%), no better than 0.8B GRPO (84%). With 100 test questions none of these gaps is significant. The honest reading is that on this small, templated task the extra parameters bought little; the cost of the larger model is in section 9.
The same map question, answered by six different trained models.
The test item below asks for the direction of Jackal's Jams relative to Giraffe's Guitars; the right answer is C, Northwest. Answers are truncated to 700 characters.
Untrained · 228 tokens · wrong
To determine the direction of **Jackal's Jams** relative to **Giraffe's Guitars**, we need to look at the positions of these two objects on the map. - **Jackal's Jams** is located in the **top-left** part of the map. - **Giraffe's Guitars** is located in the **center** of the map, slightly below the middle. So, from Giraffe's Guitars, going **up and to the left** leads to Jackal's Jams. In standard map directions: - **North** is up - **East** is right - **Southwest** is down and left - **Northeast** is up and right Since Jackal's Jams is **up and left** of Giraffe's Guitars, that corresponds to **Southwest**. Let’s check the options: A. Southwest → ✅ Correct B. Northeast → Incorrect C. [...]
SFT · 7 tokens · correct
The answer is: C. Northwest
SFT self-distilled · 182 tokens · correct
To determine the direction of **Jackal's Jams** relative to **Giraffe's Guitars**, we need to look at the positions of both objects on the map. - **Jackal's Jams** is located in the **upper-left** part of the map. - **Giraffe's Guitars** is located in the **center** of the map, slightly below the midpoint. So, from Giraffe's Guitars, going **up and to the left** would place Jackal's Jams in that direction. Let’s check the options: A. Southwest → Down and right → Incorrect B. Northeast → Up and right → Incorrect C. Northwest → Up and left → ✅ Correct D. Southeast → Down and right → Incorrect Therefore, the correct answer is **C. Northwest**. The answer is: C
GRPO · 275 tokens · correct
First, locate Giraffe's Guitars in the middle-lower area of the map. Jackal's Jams is located in the upper-left area of the map. From the perspective of looking at the map from the upper-right area (or from the viewer's viewpoint), Giraffe's Guitars serves as the reference point (origin) for the question "relative to Giraffe's Guitars." Jackal's Jams is situated horizontally to the left and vertically above Giraffe's Guitars. "Northwest" describes the location of Jackal's Jams from the viewpoint of Giraffe's Guitars, indicating that Jackal's Jams is located in the Northwest direction from Giraffe's Guitars. "Northeast" would describe a location in the Northeast direction from Giraffe's Guita [...]
RLOO · 88 tokens · correct
- Giraffe’s Guitars is in the southeast of the map. - Jackal’s Jams is to the northwest of that. - This places Jackal’s Jams in the northwest of the map. - This indicates the direction is Northwest. - Option C is Northwest. This matches the northwest of the Guitars. Therefore, the answer is C. </think> The answer is C.
GSPO · 31 tokens · correct
Looking to the northwest from Giraffe's Guitars: The answer is: C. Northwest Final line: The answer is: C. Northwest
The answer is: C. Northwest. It is the cheapest and, here, among the most accurate recipes, because the model has to read the map and answer in one step.A different test question, where it picks a letter that contradicts its own working:
Untrained · 217 tokens · wrong
To determine the direction of **Candy Land Confections** relative to **Pawn Stars**, we need to look at the spatial arrangement of the objects on the map. From the image: - **Pawn Stars** is located in the **upper right** part of the map. - **Candy Land Confections** is located in the **lower right** part of the map. So, from Pawn Stars to Candy Land Confections, you move **down and to the right**. In standard map directions: - **North** = top - **South** = bottom - **East** = right - **West** = left Since the movement is **down (south)** and **right (east)**, this corresponds to **Southeast**. Let’s check the options: A. Southwest → would be down and left B. Northeast → would be up and right C. Southeast → down and right ✅ D. Northwest → up and left The correct answer is **C. Southeast**. The answer is: C
GSPO · 29 tokens · correct
Looking to the southwest from Candy Land Confections: The answer is: A. Southwest Final line: The answer is: A. Southwest
| Run | Count questions | Accuracy | Truncated | Tokens | Predicted letters |
|---|---|---|---|---|---|
| Untrained | 39 | 8% | 41% | 461 | A: 3, B: 7, C: 12, D: 4, None: 13 |
| SFT | 39 | 85% | 0% | 8 | A: 11, B: 5, C: 13, D: 10 |
| SFT gentle | 39 | 85% | 0% | 8 | A: 9, B: 6, C: 15, D: 9 |
| SFT self-distilled | 39 | 46% | 18% | 307 | A: 6, B: 4, C: 13, D: 13, None: 3 |
| GRPO | 39 | 87% | 0% | 113 | A: 12, B: 5, C: 11, D: 11 |
| RLOO | 39 | 87% | 0% | 81 | A: 12, B: 5, C: 11, D: 11 |
| GSPO | 39 | 79% | 3% | 50 | A: 14, B: 5, C: 11, D: 9 |
Counting is where the models differ most. The untrained model almost never gets it right; after training, direction and which-object accuracy is high for most recipes, while count accuracy ranges from 46% to 87%. Whether these gains are counting or learned shortcuts is not something 100 questions can settle; the maze test below asks a related question.
| Share of identical answers | SFT | SFT gentle | SFT self-distilled | GRPO | RLOO | GSPO |
|---|---|---|---|---|---|---|
| SFT | – | 89% | 61% | 83% | 86% | 88% |
| SFT gentle | 89% | – | 62% | 82% | 89% | 89% |
| SFT self-distilled | 61% | 62% | – | 58% | 64% | 62% |
| GRPO | 83% | 82% | 58% | – | 85% | 85% |
| RLOO | 86% | 89% | 64% | 85% | – | 94% |
| GSPO | 88% | 89% | 62% | 85% | 94% | – |
Recipes with similar accuracy do not necessarily give the same answers, so the shared score hides different mistakes. This is a reason to prefer larger evaluation sets in follow-up work.
Different image type, different skill, same answer format.
We took 100 random questions from the maze-navigation task of the same dataset ("how many right turns are on the blue path?", four options, chance 25%, majority-letter baseline 32%) and evaluated every adapter without any further training. The prompt and answer line are the same.
| Run | Accuracy | Truncated | Answer line present | Tokens | Predicted letters |
|---|---|---|---|---|---|
| 0.8B Untrained | 0% | 97% | 3% | 636 | A: 1, B: 1, C: 1, null: 97 |
| 0.8B SFT | 55% | 0% | 100% | 8 | A: 5, B: 45, C: 35, D: 15 |
| 0.8B SFT gentle | 65% | 0% | 100% | 8 | A: 20, B: 26, C: 30, D: 24 |
| 0.8B SFT self-distilled | 9% | 80% | 20% | 590 | A: 6, B: 8, C: 2, D: 4, null: 80 |
| 0.8B GRPO | 60% | 4% | 96% | 274 | A: 21, B: 29, C: 28, D: 18, null: 4 |
| 0.8B RLOO | 67% | 0% | 100% | 98 | A: 27, B: 23, C: 25, D: 25 |
| 0.8B GSPO | 66% | 6% | 97% | 119 | A: 22, B: 24, C: 28, D: 23, null: 3 |
| 2B Untrained | 7% | 89% | 11% | 608 | A: 2, B: 2, C: 2, D: 5, null: 89 |
| 2B SFT | 45% | 0% | 100% | 8 | A: 49, B: 4, C: 27, D: 20 |
| 2B SFT gentle | 70% | 0% | 100% | 8 | A: 12, B: 22, C: 38, D: 28 |
| 2B GRPO | 44% | 29% | 74% | 289 | A: 18, B: 19, C: 18, D: 19, null: 26 |
Training changed how many tokens each answer costs, and with them the time per answer.
The untrained model writes 392 tokens per answer, 1006 per correct answer. SFT writes 8; GSPO 38; RLOO 87; GRPO 158. Dividing generated tokens by correct answers is the number that matters for cost: 109x fewer for SFT and 23x fewer for GSPO than the untrained model.
These are aggregate numbers for many answers in flight, so they are higher than what one user sees. Batched decode throughput was about 487 tokens/s untrained and 224-514 tokens/s after training; it was lowest for GSPO (224), whose answers are the shortest of the RL runs, probably because with short answers the evaluation is dominated by prefill and start-up rather than steady decoding (not isolated). What matters more is wall-clock time: the 100-question evaluation took 97 s untrained and 24 s after SFT.
The same 100 questions, one at a time, greedy, timed per token.
| Run | Accuracy | Prompt tokens | Gen tokens | TTFT s | Prefill tok/s | Decode tok/s | ITL p50 ms | ITL p95 ms | Total s | Peak MLX GB |
|---|---|---|---|---|---|---|---|---|---|---|
| Untrained | 42% | 304 | 349 | 0.127 | 2482 | 64 | 15.7 | 16.4 | 5.63 | 2.26 |
| SFT | 89% | 304 | 9 | 0.126 | 2496 | 71 | 15.7 | 16.2 | 0.25 | 2.09 |
| SFT gentle | 90% | 304 | 9 | 0.126 | 2497 | 72 | 15.7 | 16.1 | 0.25 | 2.09 |
| SFT self-distilled | 59% | 304 | 279 | 0.126 | 2492 | 64 | 15.7 | 16.3 | 4.51 | 2.09 |
| GRPO | 82% | 304 | 177 | 0.126 | 2491 | 64 | 15.7 | 16.4 | 2.90 | 2.09 |
| RLOO | 87% | 304 | 88 | 0.127 | 2477 | 64 | 15.7 | 16.4 | 1.50 | 2.09 |
| GSPO | 88% | 304 | 39 | 0.126 | 2482 | 66 | 15.7 | 16.2 | 0.73 | 2.09 |
Accuracy is only half of a benchmark. This chapter reads the system counters logged at every step to see what the machine did while it trained: where the time went, how busy the GPU and CPU were, and whether the laptop ever slowed down.
These are the nine LoRA runs of this study: six on Qwen3.5-0.8B and three on Qwen3.5-2B. The vision tower is frozen and its features are computed once per image, so the training steps are language-model steps over 196 image tokens plus the text; the cost of the image shows up as longer sequences, not as extra vision compute.
Across 9 runs (13 hours of training in total) the GPU was busy 69% of the time on average and the CPU only 14%: training is a GPU job, and the CPU mostly feeds it. The thermal speed limit never dropped below 100%, so the laptop never throttled, even through multi-hour runs on a plugged-in laptop.
Each RL step first samples 16 answers and then trains on them. Sampling took 68-81% of the step time (74% on average). Sampling is memory-bandwidth bound exactly as in the inference benchmark, so the lever that matters is how many tokens you generate, not how clever the optimiser is. Batching helps a lot: 16 sequences decoded together reach 175-377 tokens/s in total, against about 64 tokens/s for a single stream on the same model.
GPU utilisation sat between 31% and 87% for every run, while the CPU stayed at 9-18%.
System RAM read 13.9-14.1 GB in every run, almost the whole 16 GB, and the swap file held 4.5-9.2 GB. This is mostly macOS using memory for caches and other apps, so it says more about the machine than about training, but it also means there is no slack: the runs that peaked at 7.7 GB of MLX memory had only a few GB of headroom. The steady swap use did not visibly slow the runs.
| Run | Steps | Hours | s / step | Sampling share | Gen tok/s | Train tok/s | GPU | CPU | Peak MLX GB | RAM GB | Swap GB | Gain (pts) | Pts / hour |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 0.8B SFT | 625 | 0.5 | 2.5 | – | – | 16 | 68% | 17% | 3.7 | 14.1 | 9.2 | +50 | +101.7 |
| 0.8B SFT gentle | 250 | 0.2 | 2.4 | – | – | 16 | 72% | 18% | 3.7 | 14.0 | 8.9 | +52 | +250.8 |
| 0.8B SFT self-distilled | 230 | 0.4 | 5.5 | – | – | 236 | 87% | 12% | 7.4 | 13.9 | 4.5 | +24 | +56.5 |
| 0.8B GRPO | 625 | 3.7 | 20.7 | 72% | 374 | 334 | 86% | 14% | 7.3 | 14.1 | 7.9 | +45 | +12.1 |
| 0.8B RLOO | 625 | 2.6 | 14.4 | 68% | 377 | 384 | 82% | 11% | 7.4 | 14.0 | 5.2 | +49 | +19.0 |
| 0.8B GSPO | 625 | 1.5 | 8.4 | 75% | 376 | 206 | 87% | 10% | 7.3 | 13.9 | 4.7 | +48 | +31.5 |
| 2B SFT | 625 | 0.7 | 3.3 | – | – | 11 | 31% | 17% | 6.7 | 13.9 | 7.0 | +41 | +59.6 |
| 2B SFT gentle | 250 | 0.3 | 3.3 | – | – | 11 | 31% | 17% | 6.7 | 13.9 | 7.2 | +41 | +136.8 |
| 2B GRPO | 625 | 3.2 | 17.9 | 81% | 175 | 160 | 82% | 9% | 7.7 | 14.0 | 6.5 | +36 | +11.2 |
Gain is final minus untrained test accuracy in points; points per hour divides it by the wall-clock hours of the whole run, evaluations included (one evaluation of the test set takes about 61 s). Energy and battery use were not measured: that needs administrator access to the power counters.
| Run | s / step | of which generation | of which training | gen tok/s | train tok/s | peak MLX GB | RAM GB (max) | swap GB (max) | hours |
|---|---|---|---|---|---|---|---|---|---|
| 0.8B GRPO | 20.7 | 14.8 | 5.9 | 374 | 334 | 7.3 | 14.1 | 7.9 | 3.7 |
| 2B GRPO | 17.9 | 14.5 | 3.4 | 175 | 160 | 7.7 | 14.0 | 6.5 | 3.2 |
| 0.8B SFT | 2.5 | – | – | – | 16 | 3.7 | 14.1 | 9.2 | 0.5 |
| 0.8B SFT gentle | 2.4 | – | – | – | 16 | 3.7 | 14.0 | 8.9 | 0.2 |
| 0.8B SFT self-distilled | 5.5 | – | – | – | 236 | 7.4 | 13.9 | 4.5 | 0.4 |
| 0.8B RLOO | 14.4 | 9.7 | 4.6 | 377 | 384 | 7.4 | 14.0 | 5.2 | 2.6 |
| 0.8B GSPO | 8.4 | 6.2 | 2.1 | 376 | 206 | 7.3 | 13.9 | 4.7 | 1.5 |
| 2B SFT | 3.3 | – | – | – | 11 | 6.7 | 13.9 | 7.0 | 0.7 |
| 2B SFT gentle | 3.3 | – | – | – | 11 | 6.7 | 13.9 | 7.2 | 0.3 |
On average a 2B GRPO step took 18 s against 21 s for 0.8B, with generation 81% of the step (72% for 0.8B). The first 2B step took 43 s because untrained answers were long; steps got cheaper as the model learned to answer briefly, the same effect seen at 0.8B, and the 2B run took 3.2 h in total for 625 steps.
One seed per recipe and 100 test questions (±10 points at 95%). A single dataset slice of a templated, synthetic task. The maze-navigation results have no control (blank image, larger token cap), so transfer is suggestive only. Single-stream speed was measured on 0.8B only; the 2B figures are training and batched-evaluation numbers. We froze the vision tower; training it was not attempted.
| Run | Epoch | All | Direction | Which object | Count | Tokens | Truncated |
|---|---|---|---|---|---|---|---|
| 0.8B SFT | 0 | 39% | 61% | 57% | 8% | 392 | 29% |
| 0.8B SFT | 1 | 47% | 32% | 17% | 82% | 8 | 0% |
| 0.8B SFT | 2 | 64% | 32% | 90% | 69% | 8 | 0% |
| 0.8B SFT | 3 | 76% | 71% | 80% | 77% | 8 | 0% |
| 0.8B SFT | 4 | 86% | 90% | 87% | 82% | 8 | 0% |
| 0.8B SFT | 5 | 89% | 94% | 90% | 85% | 8 | 0% |
| 0.8B SFT gentle | 0 | 39% | 61% | 57% | 8% | 392 | 29% |
| 0.8B SFT gentle | 1 | 85% | 77% | 90% | 87% | 8 | 0% |
| 0.8B SFT gentle | 2 | 91% | 94% | 97% | 85% | 8 | 0% |
| 0.8B SFT self-distilled | 0 | 39% | 61% | 57% | 8% | 392 | 29% |
| 0.8B SFT self-distilled | 1 | 49% | 68% | 73% | 15% | 325 | 12% |
| 0.8B SFT self-distilled | 2 | 63% | 77% | 70% | 46% | 302 | 12% |
| 0.8B GRPO | 0 | 39% | 61% | 57% | 8% | 392 | 29% |
| 0.8B GRPO | 1 | 78% | 65% | 80% | 87% | 223 | 1% |
| 0.8B GRPO | 2 | 79% | 65% | 87% | 85% | 229 | 0% |
| 0.8B GRPO | 3 | 83% | 71% | 90% | 87% | 164 | 1% |
| 0.8B GRPO | 4 | 82% | 81% | 83% | 82% | 163 | 0% |
| 0.8B GRPO | 5 | 84% | 74% | 90% | 87% | 158 | 0% |
| 0.8B RLOO | 0 | 39% | 61% | 57% | 8% | 392 | 29% |
| 0.8B RLOO | 1 | 75% | 74% | 67% | 82% | 109 | 0% |
| 0.8B RLOO | 2 | 81% | 84% | 80% | 79% | 172 | 0% |
| 0.8B RLOO | 3 | 90% | 87% | 97% | 87% | 86 | 0% |
| 0.8B RLOO | 4 | 89% | 94% | 87% | 87% | 87 | 0% |
| 0.8B RLOO | 5 | 88% | 90% | 87% | 87% | 87 | 0% |
| 0.8B GSPO | 0 | 39% | 61% | 57% | 8% | 392 | 29% |
| 0.8B GSPO | 1 | 84% | 81% | 83% | 87% | 57 | 0% |
| 0.8B GSPO | 2 | 84% | 84% | 90% | 79% | 34 | 0% |
| 0.8B GSPO | 3 | 89% | 97% | 87% | 85% | 34 | 0% |
| 0.8B GSPO | 4 | 88% | 94% | 90% | 82% | 38 | 1% |
| 0.8B GSPO | 5 | 87% | 94% | 90% | 79% | 38 | 1% |
| 2B SFT | 0 | 49% | 74% | 77% | 8% | 322 | 14% |
| 2B SFT | 1 | 76% | 58% | 83% | 85% | 8 | 0% |
| 2B SFT | 2 | 83% | 71% | 97% | 82% | 8 | 0% |
| 2B SFT | 3 | 88% | 97% | 90% | 79% | 8 | 0% |
| 2B SFT | 4 | 91% | 100% | 90% | 85% | 8 | 0% |
| 2B SFT | 5 | 90% | 100% | 90% | 82% | 8 | 0% |
| 2B SFT gentle | 0 | 49% | 74% | 77% | 8% | 322 | 14% |
| 2B SFT gentle | 1 | 87% | 84% | 97% | 82% | 8 | 0% |
| 2B SFT gentle | 2 | 90% | 97% | 93% | 82% | 8 | 0% |
| 2B GRPO | 0 | 49% | 74% | 77% | 8% | 322 | 14% |
| 2B GRPO | 1 | 89% | 90% | 97% | 82% | 42 | 0% |
| 2B GRPO | 2 | 89% | 97% | 87% | 85% | 29 | 0% |
| 2B GRPO | 3 | 86% | 94% | 80% | 85% | 33 | 0% |
| 2B GRPO | 4 | 85% | 94% | 80% | 82% | 49 | 0% |
| 2B GRPO | 5 | 85% | 90% | 83% | 82% | 53 | 0% |
macbook-posttrain-spatial. Metrics, rollouts and evaluation answers are kept under runs_spatial/; adapter weights were not committed.Copyright © 2022-2026 Azmine Toushik Wasi. All rights reserved.