Simplify model card to usage

README.md

@@ -11,104 +11,21 @@ tags:
   - math
 ---
 
-# Ouro-1.4B-Thinking — depth-as-action GRPO (step 400)
-
-RL fine-tune of [ByteDance/Ouro-1.4B-Thinking](https://huggingface.co/ByteDance/Ouro-1.4B-Thinking)
-in which the **loop count `r` is part of the RL action and is priced**. Trained with GRPO for 400
-steps, rolling out every prompt at `r ∈ {2, 4, 8, 16}` with advantages computed *across* depth
-groups and a linear cost `λ·(r/16)`, `λ = 0.1`.
-
-**Depths `r = 24` and `r = 32` were never rolled out during RL.** They are evaluation-only.
-
-Initialized from the SFT baseline
-[`omar81939/Ouro-1.4B-Thinking-depth-SFT`](https://huggingface.co/omar81939/Ouro-1.4B-Thinking-depth-SFT) (called **W-A** below) — you need that checkpoint to
-reproduce the comparison, because every number here is GRPO-vs-W-A, not GRPO-vs-base-Ouro.
-
-## What this checkpoint does and does not do
-
-**Read this before quoting a number.**
-
-- It substantially repairs the base model's collapse at loop depths beyond the trained horizon.
-- It does **not** make depth buy accuracy. This model's own accuracy still *falls* monotonically
-  with `r`: 0.840 → 0.802 → 0.660 → 0.478 across `r = 4/16/24/32`. Deeper is still worse.
-- The mechanism is largely **termination**, not reasoning. At `r = 32` the W-A baseline hits the
-  3,072-token budget on 84.4% of MATH-500 problems versus 17.2% for this model, and of the 168
-  problems this model gets right where W-A fails, 160 are cases where W-A ran out of tokens or
-  never emitted a boxed answer.
-- It is a **single training seed** (1234) evaluated on **MATH-500 only**. A confirmation run on a
-  second seed and an objective ablation (`λ = 0`, within-depth advantages) were pre-registered and
-  had not run when this was uploaded.
-- The model also carries a depth-selection head trained against `Q(r) = mean reward − λ·(r/16)`,
-  but the sampler never consulted it during training and its behavior was never evaluated. **No
-  adaptive per-prompt depth allocation is claimed.**
-
-## Results
-
-MATH-500, greedy, 3,072-token budget, n = 500. `trunc` = share of generations that hit the budget.
-
-| depth | base Ouro | W-A (SFT) acc / trunc | this model acc / trunc | Δ acc |
-|---|---|---|---|---|
-| r = 4  | 0.710 | 0.676 / 34.6% | **0.840** / 6.2%  | +16.4 |
-| r = 16 | 0.058 | 0.620 / 38.8% | **0.802** / 5.6%  | +18.2 |
-| r = 24 (never trained) | 0.004 | 0.378 / 62.0% | **0.660** / 8.6%  | +28.2 |
-| r = 32 (never trained) | 0.000 | 0.178 / 84.4% | **0.478** / 17.2% | +30.0 |
-
-A pre-registered gate — the falloff from `r = 16` must be shallower than the baseline's, judged on
-a paired item-level bootstrap (20,000 draws, Bonferroni-adjusted 97.5% lower bounds) — passed at
-both untrained depths: `W(24) = +10.0` (LB +3.8) and `W(32) = +11.8` (LB +4.8). The identical
-analysis **failed** 50 steps earlier at step 350 (LBs −1.0 and −0.8), which is why the
-single-seed caveat above matters.
-
-## Usage
-
-The loop count is a config field, `total_ut_steps`. It is read at model-construction time, so set
-it when you load:
+# Usage
 
 ```python
 from transformers import AutoModelForCausalLM, AutoTokenizer
 
 REPO = "omar81939/Ouro-1.4B-Thinking-depth-GRPO"
-tok = AutoTokenizer.from_pretrained(REPO)
+DEPTH = 16
+
+tokenizer = AutoTokenizer.from_pretrained(REPO)
 model = AutoModelForCausalLM.from_pretrained(
-    REPO, trust_remote_code=True, dtype="bfloat16",
-    total_ut_steps=16,          # <- the dial: 2, 4, 8, 16 were trained; 24/32 are out of support
+    REPO,
+    trust_remote_code=True,
+    dtype="bfloat16",
+    total_ut_steps=DEPTH,
 )
 ```
 
-With vLLM, pass it as an engine override (`hf_overrides={"total_ut_steps": r}`) — it cannot be
-changed per-request after engine init.
-
-The exported weights are BF16 and the config ships `total_ut_steps: 4`, matching the base model's
-default. Evaluation numbers above were produced by overriding it per run.
-
-## Training
-
-| | |
-|---|---|
-| Base | ByteDance/Ouro-1.4B-Thinking → SFT (W-A) → GRPO |
-| RL steps | 400 |
-| Rollouts | 48 prompts × 4 depths × 4 generations = 768 sequences/step |
-| Depths sampled | r ∈ {2, 4, 8, 16} |
-| Reward | first-boxed-answer correctness, + a small bonus for emitting a parseable box |
-| Depth pricing | `Q(r) = mean reward − λ·(r/16)`, λ = 0.1, advantages across depth groups |
-| Prompts | DeepScaleR, decontaminated against every eval set, pass@8-filtered → 8,932 prompts |
-| Hardware | 8 × H100, ~23 days wall clock on preemptible partitions |
-| Sampling (train) | temperature 1.0, top-p 1.0, 2,048-token responses |
-
-The run was bitwise deterministic: every step logged parity records (307,200 sequence-level
-comparisons across the run, zero failures), roughly 40 preemptions all replayed byte-identically,
-and a mid-run migration reproduced a step byte-for-byte on a different cluster on another
-continent.
-
-## Limitations
-
-Single seed. One benchmark. Math only. English only. The base model's own card notes it is a
-research artifact, not a production model, and that applies here at least as strongly. Deeper loop
-counts cost proportionally more compute for *lower* accuracy on this checkpoint, so there is no
-setting of `r` above 4 that is recommended for use — the interest here is scientific.
-
-## License and attribution
-
-Apache-2.0, inherited from the base model. Base weights © ByteDance Seed, from
-[Ouro-1.4B-Thinking](https://huggingface.co/ByteDance/Ouro-1.4B-Thinking) — see
-["Scaling Latent Reasoning via Looped Language Models"](https://arxiv.org/abs/2510.25741).
+With vLLM, set `hf_overrides={"total_ut_steps": DEPTH}` when creating the engine.