File size: 2,930 Bytes
b6d5136 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 | ---
base_model: Qwen/Qwen3-1.7B
library_name: peft
tags: [opsd, lora, math-reasoning, aime]
---
# OPSD LoRA checkpoints (Qwen3-1.7B)
On-Policy Self-Distillation adapters on **Qwen/Qwen3-1.7B** (LoRA r=64, α=128, targets q/k/v/o/gate/up/down_proj, 100 steps).
Each `checkpoints/**/` leaf is a standalone PEFT adapter. Recipes in `configs/`.
## Load
```python
from transformers import AutoModelForCausalLM
from peft import PeftModel
base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-1.7B", torch_dtype="bfloat16")
model = PeftModel.from_pretrained(base, "hbin0701/opsd-lora", subfolder="checkpoints/thinkablate/open/run1")
```
(vLLM: serve Qwen/Qwen3-1.7B with `--enable-lora` and point at a downloaded adapter subfolder.)
## Checkpoints (AIME24 Avg@12, val_n=12, thinking eval)
| path | config | AIME24 |
|---|---|---|
| `checkpoints/thinkablate/nothink/run1` | direct ref teacher, teacher_thinking=False | 54.72 |
| `checkpoints/thinkablate/nothink/run2` | direct ref teacher, teacher_thinking=False | 55.83 |
| `checkpoints/thinkablate/nothink/run3` | direct ref teacher, teacher_thinking=False | 53.06 |
| `checkpoints/thinkablate/noopen/run1` | teacher_thinking=True, no open block | 55.28 |
| `checkpoints/thinkablate/noopen/run2` | teacher_thinking=True, no open block | 54.72 |
| `checkpoints/thinkablate/noopen/run3` | teacher_thinking=True, no open block | 53.89 |
| `checkpoints/thinkablate/open/run1` | teacher_thinking=True + teacher_open_think | 51.39 |
| `checkpoints/thinkablate/open/run2` | teacher_thinking=True + teacher_open_think | 55.56 |
| `checkpoints/thinkablate/open/run3` | teacher_thinking=True + teacher_open_think | 53.06 |
| `checkpoints/reference/nothink` | raw reference as teacher ctx, no thinking (ref_nothink) | 55.83 |
| `checkpoints/reference/think/run1` | reason_first + teacher_thinking_on | 49.72 |
| `checkpoints/reference/think/run2` | reason_first + teacher_thinking_on | 53.61 |
| `checkpoints/reference/think/run3` | reason_first + teacher_thinking_on | 48.89 |
| `checkpoints/reference/think/run4` | reason_first + teacher_thinking_on | 49.17 |
| `checkpoints/trd_yr` | TRD forward-KL on refined target y_r | 39.72 |
| `checkpoints/maker/p1_surgical` | maker-prefix variant p1_surgical | 48.61 |
| `checkpoints/maker/p2_graduated` | maker-prefix variant p2_graduated | 47.50 |
| `checkpoints/maker/p2b_flywheel` | maker-prefix variant p2b_flywheel | 50.83 |
| `checkpoints/maker/maker_nothink` | maker prefix, non-thinking teacher | — |
| `checkpoints/maker/dynmaker_thinkOFF_run1` | dynamic 35B maker, thinkOFF | — |
Notes: `nothink/noopen/open` are the 3-seed teacher-style ablation (all ≈54 within noise). `reference/nothink`=55.83 (n=1). `trd_yr`=forward-KL on the refined target. `configs/variants/` holds the maker-variant definitions; `configs/flags.md` the flag combos. Only the final (step-100) checkpoint is included; intermediate 25/50/75 exist upstream.
|