Mtrini-SVL-2.0 round 7, Gated Self-Training with DPO (5 steps)
Browse files- README.md +7 -5
- eval_results.json +30 -10
- model.safetensors +1 -1
- training_log.json +1 -1
README.md
CHANGED
|
@@ -18,11 +18,11 @@ Quick checks run on the base model and on this model with the same questions and
|
|
| 18 |
|
| 19 |
| | Qwen3.5-9B | Before | After |
|
| 20 |
|---|---|---|---|
|
| 21 |
-
| AIME 2025 (30 problems, avg of
|
| 22 |
-
|
|
| 23 |
-
|
|
| 24 |
-
| DarijaMMLU (400 questions) | 34.5% |
|
| 25 |
-
| Identity (says it's Mtrini by Compiwer AI) | 0% |
|
| 26 |
|
| 27 |
## Training
|
| 28 |
|
|
@@ -34,6 +34,8 @@ Three rounds of LoRA SFT, each merged before the next:
|
|
| 34 |
|
| 35 |
4. Round 4 (123 steps × 16 examples): a new slice of OpenThoughts3 math, code and science reasoning, more Darija and identity answers.
|
| 36 |
|
|
|
|
|
|
|
| 37 |
## Usage
|
| 38 |
|
| 39 |
Use it like Qwen3.5-9B. Thinking is on by default; pass `enable_thinking=False` to the chat template for direct answers.
|
|
|
|
| 18 |
|
| 19 |
| | Qwen3.5-9B | Before | After |
|
| 20 |
|---|---|---|---|
|
| 21 |
+
| AIME 2025 (30 problems, avg of 8 tries) | – | 62.1% | 61.7% |
|
| 22 |
+
| LiveCodeBench v6 (2025 problems, pass@1) | – | 47.5% | 48.8% |
|
| 23 |
+
| GPQA Diamond (pass@1) | – | 60.6% | 62.1% |
|
| 24 |
+
| DarijaMMLU (400 questions) | 34.5% | 64.0% | 64.2% |
|
| 25 |
+
| Identity (says it's Mtrini by Compiwer AI) | 0.0% | 93.8% | 93.8% |
|
| 26 |
|
| 27 |
## Training
|
| 28 |
|
|
|
|
| 34 |
|
| 35 |
4. Round 4 (123 steps × 16 examples): a new slice of OpenThoughts3 math, code and science reasoning, more Darija and identity answers.
|
| 36 |
|
| 37 |
+
5. Round 7, Gated Self-Training with preference pairs (5 DPO steps): on hard math, coding problems checked by their tests, and hard science, its own right answers preferred over its own wrong ones of similar length, plus identity pairs.
|
| 38 |
+
|
| 39 |
## Usage
|
| 40 |
|
| 41 |
Use it like Qwen3.5-9B. Thinking is on by default; pass `enable_thinking=False` to the chat template for direct answers.
|
eval_results.json
CHANGED
|
@@ -6,17 +6,37 @@
|
|
| 6 |
"humaneval": 0.9390243902439024
|
| 7 |
},
|
| 8 |
"previous": {
|
| 9 |
-
"
|
| 10 |
-
"
|
| 11 |
-
"
|
| 12 |
-
"
|
| 13 |
-
"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 14 |
},
|
| 15 |
"svl20": {
|
| 16 |
-
"
|
| 17 |
-
"
|
| 18 |
-
"
|
| 19 |
-
"
|
| 20 |
-
"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 21 |
}
|
| 22 |
}
|
|
|
|
| 6 |
"humaneval": 0.9390243902439024
|
| 7 |
},
|
| 8 |
"previous": {
|
| 9 |
+
"livecodebench": 0.475,
|
| 10 |
+
"livecodebench/easy": 0.9523809523809523,
|
| 11 |
+
"livecodebench/hard": 0.16666666666666666,
|
| 12 |
+
"livecodebench/medium": 0.5217391304347826,
|
| 13 |
+
"gpqa_diamond": 0.6060606060606061,
|
| 14 |
+
"gpqa/biology": 0.42105263157894735,
|
| 15 |
+
"gpqa/chemistry": 0.41935483870967744,
|
| 16 |
+
"gpqa/physics": 0.8488372093023255,
|
| 17 |
+
"aime2025": 0.6208333333333333,
|
| 18 |
+
"aime/algebra": 0.7222222222222222,
|
| 19 |
+
"aime/combinatorics": 0.5416666666666666,
|
| 20 |
+
"aime/geometry": 0.6607142857142857,
|
| 21 |
+
"aime/number theory": 0.525,
|
| 22 |
+
"darija_mmlu": 0.64,
|
| 23 |
+
"identity": 0.9375
|
| 24 |
},
|
| 25 |
"svl20": {
|
| 26 |
+
"livecodebench": 0.4875,
|
| 27 |
+
"livecodebench/easy": 0.9523809523809523,
|
| 28 |
+
"livecodebench/hard": 0.1388888888888889,
|
| 29 |
+
"livecodebench/medium": 0.6086956521739131,
|
| 30 |
+
"gpqa_diamond": 0.6212121212121212,
|
| 31 |
+
"gpqa/biology": 0.5263157894736842,
|
| 32 |
+
"gpqa/chemistry": 0.3870967741935484,
|
| 33 |
+
"gpqa/physics": 0.8953488372093024,
|
| 34 |
+
"aime2025": 0.6166666666666667,
|
| 35 |
+
"aime/algebra": 0.5972222222222222,
|
| 36 |
+
"aime/combinatorics": 0.5277777777777778,
|
| 37 |
+
"aime/geometry": 0.8392857142857143,
|
| 38 |
+
"aime/number theory": 0.5,
|
| 39 |
+
"darija_mmlu": 0.6425,
|
| 40 |
+
"identity": 0.9375
|
| 41 |
}
|
| 42 |
}
|
model.safetensors
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
size 18819722392
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:55e74288b86ccf17add3fb570260061fba330c100526392440057cdcd24e815f
|
| 3 |
size 18819722392
|
training_log.json
CHANGED
|
@@ -1 +1 @@
|
|
| 1 |
-
{"steps":
|
|
|
|
| 1 |
+
{"steps": 5, "max_steps": 5, "log": [{"loss": 0.6931473016738892, "grad_norm": 17.957149505615234, "learning_rate": 1.6666666666666667e-06, "entropy": 0.6042060684412718, "num_tokens": 129113.0, "logits/chosen": -2.031570067427444, "logits/rejected": -2.066084378741378, "mean_token_accuracy": 0.8573183789849281, "rewards/chosen": 0.0, "rewards/rejected": 0.0, "rewards/accuracies": 0.0, "rewards/margins": 0.0, "logps/chosen": -1720.7409473061562, "logps/rejected": -1911.8703248500824, "epoch": 0.43243243243243246, "step": 2}, {"loss": 0.7156742215156555, "grad_norm": 14.256765365600586, "learning_rate": 5e-06, "entropy": 0.7723287399858236, "num_tokens": 235146.0, "logits/chosen": -1.9302834164158855, "logits/rejected": -1.978542090787214, "mean_token_accuracy": 0.832091361284256, "rewards/chosen": -0.020962871101801284, "rewards/rejected": 0.020634264779801015, "rewards/accuracies": 0.4375, "rewards/margins": -0.04159713693661615, "logps/chosen": -1395.894810795784, "logps/rejected": -1786.6366989016533, "epoch": 0.8648648648648649, "step": 4}, {"train_runtime": 249.6526, "train_samples_per_second": 0.16, "train_steps_per_second": 0.02, "total_flos": 1.5657738870199104e+16, "train_loss": 0.7078817367553711, "entropy": 0.6014999866485595, "num_tokens": 277106.0, "logits/chosen": -2.0039181554060184, "logits/rejected": -1.9878743479109093, "mean_token_accuracy": 0.8745649218559265, "rewards/chosen": -0.044874840695410964, "rewards/rejected": -0.00777633311226964, "rewards/accuracies": 0.6, "rewards/margins": -0.03709850497543812, "logps/chosen": -1500.4318752288818, "logps/rejected": -1871.2910789489747, "epoch": 1.0, "step": 5}]}
|