Downloads · 30 days
12
16% of all-time downloads
kzhao5/opd-rethink-k2-horizon-7b-semantic-class-step200
opd-rethink-k2-horizon-7b-semantic-class-step200 is a machine learning model from kzhao5. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as other.
Step-200 student checkpoint from the OPDRethink EOS-alignment study, K2-Horizon extension. The same run was later extended to 400 steps; checkpoints 220…400 are included under checkpoints/ (see the per-step table).
Downloads · 30 days
12
16% of all-time downloads
All-time downloads
77
Public
Parameters
9B
360 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors378 GB · 100%
From the Hugging Face model README
Step-200 student checkpoint from the OPD_Rethink EOS-alignment study, K2-Horizon extension. The same run was later extended to 400 steps; checkpoints 220…400 are included under checkpoints/ (see the per-step table).
| Student (init) | IFM/K2-Horizon-7B revision pretrain_final (pretraining-final, sha fdbc38e9…) |
| Teacher | IFM/K2-Horizon-7B revision main (final post-trained checkpoint, sha 586b03f0…; byte-identical to sft_2_final) |
| Method | on-policy distillation (verl, token-level student−teacher log-prob reward), 200 steps, batch 16 × 4 rollouts, prompt 1024 + response ≤ 7168 tokens (student native context 8192) |
| EOS mode | semantic_class: EOD `< |
| Prompt format | teacher's chat template with an empty think block forced (enable_thinking=False, see chat_template.jinja); the official thinking-by-default template is kept as chat_template.official.jinja |
| Tokenizer | main's tokenizer files (token ids/merges are identical across revisions; only special-token surface strings were renamed) |
| wandb | https://wandb.ai/shawn2002/OPD_Rethink/runs/74a619fc |
Official K2HorizonForCausalLM layout (grouped RMSNorm, layernorm_num_groups=4) with the official remote code,
so it loads like the upstream model (transformers ≥ 5.13):
from transformers import AutoModelForCausalLM, AutoTokenizer
m = AutoModelForCausalLM.from_pretrained(REPO, trust_remote_code=True, dtype="bfloat16")
tok = AutoTokenizer.from_pretrained(REPO)
Stop on ids [1, 250019] (generation_config.json). Export was verified against the training-side
implementation (logits of the exported model under the official code vs the training checkpoint under our
patched Llama implementation) before upload; see deliverables/k2/EQUIVALENCE_REPORT.md in the project.
Unlike Llama-3.2 / Gemma-3 / Qwen3 students, semantic_class does not stop the length inflation here:
mean response length 1361 → ~6000 tokens, clip ratio 0.09 → 0.7–0.8, share of rollouts ending with a stop
token 0.91 → ~0.2 over 200 steps. The teacher assigns only 0.02–0.09 probability to STOP at the student's
endings (it wants a long reasoning trace first), so the sampled-EOS advantage is already ≈ −4 at step 1 —
a when-to-stop mismatch rather than a which-EOS-token mismatch. The vanilla-OPD control of the same
pair tracks the same trajectory. Per-checkpoint evaluation (AIME24/25, AMC23, 16 rollouts, max_tokens 7168) is tabulated below; raw JSON/JSONL is in the public
dataset repo kzhao5/opd-rethink-eval-results under opd-rethink-20260905T180000Z/k2_semantic_class/.
License: inherits the base model's license.
All 10 training checkpoints are included: the repo root holds step 200, and checkpoints/step_0020 … checkpoints/step_0200
hold every evaluated checkpoint (same HF format, tokenizer included in each folder).
from transformers import AutoModelForCausalLM
m = AutoModelForCausalLM.from_pretrained("kzhao5/opd-rethink-k2-horizon-7b-semantic-class-step200", subfolder="checkpoints/step_0100", trust_remote_code=True, dtype="bfloat16")
Per-checkpoint evaluation (16 rollouts/question, T=0.7, top_p=0.95, max_tokens 7168; mean accuracy / truncation rate):
| step | AIME24 acc / trunc | AIME25 acc / trunc | AMC23 acc / trunc | avg len (AMC23) |
|---|---|---|---|---|
| 20 | 0.006 / 0.41 | 0.004 / 0.47 | 0.042 / 0.31 | 2429 |
| 40 | 0.019 / 0.62 | 0.021 / 0.68 | 0.163 / 0.45 | 3644 |
| 60 | 0.027 / 0.70 | 0.023 / 0.75 | 0.191 / 0.50 | 4091 |
| 80 | 0.040 / 0.81 | 0.025 / 0.82 | 0.217 / 0.55 | 4300 |
| 100 | 0.035 / 0.87 | 0.046 / 0.85 | 0.264 / 0.58 | 4632 |
| 120 | 0.042 / 0.88 | 0.052 / 0.90 | 0.272 / 0.60 | 4756 |
| 140 | 0.033 / 0.87 | 0.052 / 0.88 | 0.258 / 0.62 | 4841 |
| 160 | 0.042 / 0.88 | 0.065 / 0.87 | 0.273 / 0.61 | 4840 |
| 180 | 0.042 / 0.91 | 0.058 / 0.89 | 0.272 / 0.61 | 4883 |
| 200 | 0.035 / 0.94 | 0.058 / 0.91 | 0.294 / 0.60 | 4881 |
| 220 | 0.044 / 0.94 | 0.058 / 0.94 | 0.287 / 0.70 | 5370 |
| 240 | 0.025 / 0.96 | 0.048 / 0.95 | 0.264 / 0.75 | 5592 |
| 260 | 0.040 / 0.96 | 0.054 / 0.96 | 0.286 / 0.72 | 5459 |
| 280 | 0.050 / 0.96 | 0.065 / 0.95 | 0.294 / 0.76 | 5717 |
| 300 | 0.056 / 0.97 | 0.056 / 0.95 | 0.286 / 0.76 | 5670 |
| 320 | 0.033 / 0.96 | 0.056 / 0.96 | 0.291 / 0.77 | 5734 |
| 340 | 0.050 / 0.96 | 0.065 / 0.96 | 0.311 / 0.77 | 5743 |
| 360 | 0.046 / 0.97 | 0.075 / 0.96 | 0.311 / 0.75 | 5628 |
| 380 | 0.056 / 0.98 | 0.062 / 0.97 | 0.292 / 0.82 | 6037 |
| 400 | 0.056 / 0.98 | 0.060 / 0.97 | 0.256 / 0.81 | 6036 |