| --- |
| 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. |
| |
