huggingface · qgallouedec · Sep 17, 2026 · Sep 9, 2026 · Sep 10, 2026 · Sep 17, 2026
diff --git a/docs/source/_toctree.yml b/docs/source/_toctree.yml
@@ -107,8 +107,6 @@
     title: Asynchronous GRPO
   - local: bema_for_reference_model
     title: BEMA for Reference Model
-  - local: bco_trainer
-    title: BCO
   - local: cpo_trainer
     title: CPO
   - local: gkd_trainer

diff --git a/docs/source/bco_trainer.md b/docs/source/bco_trainer.md
diff --git a/docs/source/dataset_formats.md b/docs/source/dataset_formats.md
@@ -412,7 +412,6 @@ Choosing the right dataset type depends on the task you are working on and the s
 | [`RewardTrainer`] | [Preference (implicit prompt recommended)](#preference) |
 | [`RLOOTrainer`] | [Prompt-only](#prompt-only) |
 | [`SFTTrainer`] | [Language modeling](#language-modeling) or [Prompt-completion](#prompt-completion) |
-| [`experimental.bco.BCOTrainer`] | [Unpaired preference](#unpaired-preference) or [Preference (explicit prompt recommended)](#preference) |
 | [`experimental.cpo.CPOTrainer`] | [Preference (explicit prompt recommended)](#preference) |
 | [`experimental.gkd.GKDTrainer`] | [Prompt-completion](#prompt-completion) |
 | [`experimental.nash_md.NashMDTrainer`] | [Prompt-only](#prompt-only) |

diff --git a/docs/source/dpo_trainer.md b/docs/source/dpo_trainer.md
@@ -114,7 +114,7 @@ Several formulations of the objective have been proposed in the literature. Init
 | `"exo_pair"` | The [EXO](https://huggingface.co/papers/2402.00856) authors propose reverse-KL preference optimization. `label_smoothing` must be strictly greater than `0.0`; a recommended value is `1e-3` (see Eq. 16 for the simplified pairwise variant). The full method uses `K>2` SFT completions and approaches PPO as `K` grows. |
 | `"nca_pair"` | The [NCA](https://huggingface.co/papers/2402.05369) authors shows that NCA optimizes the absolute likelihood for each response rather than the relative likelihood. |
 | `"robust"` | The [Robust DPO](https://huggingface.co/papers/2403.00409) authors propose an unbiased DPO loss under noisy preferences. Use `label_smoothing` in [`DPOConfig`] to model label-flip probability; valid values are in the range `[0.0, 0.5)`. |
-| `"bco_pair"` | The [BCO](https://huggingface.co/papers/2404.04656) authors train a binary classifier whose logit serves as a reward so that the classifier maps {prompt, chosen completion} pairs to 1 and {prompt, rejected completion} pairs to 0. For unpaired data, we recommend the dedicated [`experimental.bco.BCOTrainer`]. |
+| `"bco_pair"` | The [BCO](https://huggingface.co/papers/2404.04656) authors train a binary classifier whose logit serves as a reward so that the classifier maps {prompt, chosen completion} pairs to 1 and {prompt, rejected completion} pairs to 0. |
 | `"sppo_hard"` | The [SPPO](https://huggingface.co/papers/2405.00675) authors claim that SPPO is capable of solving the Nash equilibrium iteratively by pushing the chosen rewards to be as large as 1/2 and the rejected rewards to be as small as -1/2 and can alleviate data sparsity issues. The implementation approximates this algorithm by employing hard label probabilities, assigning 1 to the winner and 0 to the loser. |
 | `"aot"`  or `loss_type="aot_unpaired"` | The [AOT](https://huggingface.co/papers/2406.05882) authors propose Distributional Preference Alignment via Optimal Transport. `loss_type="aot"` is for paired data; `loss_type="aot_unpaired"` is for unpaired data. Both enforce stochastic dominance via sorted quantiles; larger per-GPU batch sizes help. |
 | `"apo_zero"` or `loss_type="apo_down"` | The [APO](https://huggingface.co/papers/2408.06266) method introduces an anchored objective. `apo_zero` boosts winners and downweights losers (useful when the model underperforms the winners). `apo_down` downweights both, with stronger pressure on losers (useful when the model already outperforms winners). |

diff --git a/docs/source/index.md b/docs/source/index.md
@@ -73,7 +73,6 @@ Below is an overview of TRL trainers, organized by maturity and method type.
 #### Offline methods
 
 - [BEMA for Reference Model](bema_for_reference_model)
-- [`BCOTrainer`](bco_trainer)
 - [`CPOTrainer`](cpo_trainer)
 - [`ORPOTrainer`](orpo_trainer)
 - [`TPOTrainer`](tpo_trainer)

diff --git a/docs/source/paper_index.md b/docs/source/paper_index.md
@@ -954,8 +954,6 @@ training_args = DPOConfig(
 )
 ```
 
-For the unpaired version, the user should utilize [`experimental.bco.BCOConfig`] and [`experimental.bco.BCOTrainer`].
-
 ### Learn Your Reference Model for Real Good Alignment
 
 **📜 Paper**: https://huggingface.co/papers/2404.09656

diff --git a/pyproject.toml b/pyproject.toml
@@ -44,10 +44,6 @@ Homepage = "https://github.com/huggingface/trl"
 trl = "trl.cli:main"
 
 [project.optional-dependencies]
-bco = [
-    "scikit-learn",
-    "joblib"
-]
 deepspeed = [
     "deepspeed>=0.18.6",
     "transformers!=5.1.0",  # see transformers#43780
@@ -69,9 +65,6 @@ quality = [
 quantization = [
     "bitsandbytes"
 ]
-scikit = [
-    "scikit-learn"
-]
 test = [
     "pytest-cov",
     "pytest-datadir>=1.7.0",  # lazy datadirs
@@ -101,9 +94,6 @@ harbor = [
     "harbor>=0.13.0; python_version >= '3.12'",  # harbor requires Python 3.12+ (pulls its sandbox backends)
 ]
 dev = [
-    # bco
-    "scikit-learn",
-    "joblib",
     # deepspeed
     "deepspeed>=0.18.6",
     # kernels: transformers renamed the "hub-kernels" extra to "kernels" in 5.1.0
@@ -120,7 +110,6 @@ dev = [
     "hf-doc-builder",
     # quantization
     "bitsandbytes",
-    # scikit: included in bco
     # test
     "pytest-cov",
     "pytest-datadir>=1.7.0",  # lazy datadirs
-Original file line number
+Diff line change
@@ Expand Up / @@ -954,8 +954,6 @@ training_args = DPOConfig( @@
     )
     ```
-    For the unpaired version, the user should utilize [`experimental.bco.BCOConfig`] and [`experimental.bco.BCOTrainer`].
     ### Learn Your Reference Model for Real Good Alignment
     **📜 Paper**: https://huggingface.co/papers/2404.09656
@@ Expand Down @@