judge-eval 0.1.0__tar.gz

judge_eval-0.1.0/LICENSE

@@ -0,0 +1,201 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

judge_eval-0.1.0/PKG-INFO

@@ -0,0 +1,14 @@
+Metadata-Version: 2.4
+Name: judge-eval
+Version: 0.1.0
+Requires-Python: >=3.10
+License-File: LICENSE
+Requires-Dist: ray[default]>=2.9
+Requires-Dist: vllm>=0.4
+Requires-Dist: datasets
+Requires-Dist: transformers
+Requires-Dist: jinja2
+Requires-Dist: hydra-core>=1.3
+Provides-Extra: dev
+Requires-Dist: pytest; extra == "dev"
+Dynamic: license-file

judge_eval-0.1.0/README.md

@@ -0,0 +1,416 @@
+# judge-eval-hub
+
+[![License: Apache 2.0](https://img.shields.io/badge/License-Apache%202.0-blue.svg)](https://opensource.org/licenses/Apache-2.0)
+[![Hydra](https://img.shields.io/badge/Config-Hydra-89b4fa)](https://hydra.cc/)
+
+Unified judge evaluation hub — supports **generative** and **discriminative** judge/reward models, **pairwise** and **pointwise** input types, across multiple datasets in parallel.
+
+Loads the model **once**, evaluates across **multiple datasets simultaneously**, and maximises GPU utilisation via a Ray-based pipeline with vLLM.
+
+**Key advantages:**
+- **Easy to extend** — adding a new dataset is just a YAML spec + adapter function. Adding a new template is just a Python class + registry entry. No pipeline changes needed.
+- **Unified pipeline** — same architecture for generative and discriminative judge/reward models, both powered by vLLM with tensor parallelism.
+- **Multi-dataset parallel** — evaluate one model across all benchmarks in a single run.
+- **Dataset-specific aggregation** — each dataset defines its own reduction and aggregation strategy (e.g., micro, macro, weighted sections).
+- **Resume support** — checkpoint per unit, re-run to pick up where you left off.
+
+---
+
+## Quick Start
+
+```bash
+# Install
+pip install -e .
+
+# Generative judge/reward model (default config: hub_config.yaml)
+judge-eval
+
+# Discriminative judge/reward model
+judge-eval --config-name discriminative
+
+# CLI overrides (Hydra)
+judge-eval model.path=Qwen/Qwen3-8B-Instruct
+judge-eval --config-name discriminative model.path=Skywork/Skywork-Reward-V2-Qwen3-8B
+judge-eval model.engine.gpu_memory_utilization=0.95
+judge-eval "resources.gpu_groups=[[0,1],[2,3]]"
+
+# Validate config — resolves datasets/templates, prints summary, no model loading
+judge-eval run.dry_run=true
+
+# Preview formatted prompts — 1 sample per dataset, no inference
+judge-eval run.show_prompt=true
+```
+
+---
+
+## Supported Datasets
+
+| Dataset                  | Adapter          | Unit Structure | Reduction   | Overall Aggregation        |
+|--------------------------|------------------|----------------|-------------|----------------------------|
+| `nvidia/HelpSteer2`      | `helpsteer2`     | pair           | direct      | micro                      |
+| `nvidia/HelpSteer3`      | `helpsteer3`     | pair           | direct      | micro                      |
+| `Anthropic/hh-rlhf`      | `hh_rlhf`        | pair           | direct      | micro                      |
+| `allenai/reward-bench`   | `reward_bench`   | pair           | direct      | custom (weighted sections) |
+| `allenai/reward-bench-2` | `reward_bench_2` | one_vs_many    | all_correct | macro:subset               |
+| `THU-KEG/RM-Bench`       | `rm_bench`       | pair           | direct      | macro:domain               |
+| `ScalerLab/JudgeBench`   | `judgebench`     | pair           | direct      | micro                      |
+
+## Supported Templates & Parsers
+
+| Template Name              | Input Type | Output Format            | Parser                  |
+|----------------------------|------------|--------------------------|-------------------------|
+| `pairwise-fixed`           | Pairwise   | `<verdict>A/B</verdict>` | `parse_binary_verdict`  |
+| `pairwise-fixed-concrete`  | Pairwise   | `<verdict>A/B</verdict>` | `parse_binary_verdict`  |
+| `pairwise-adaptive`        | Pairwise   | `<verdict>A/B</verdict>` | `parse_binary_verdict`  |
+| `pairwise-adaptive-verify` | Pairwise   | `<verdict>A/B</verdict>` | `parse_binary_verdict`  |
+| `pairwise-openrubric`      | Pairwise   | `Winner: Response A/B`   | `parse_binary_winner`   |
+| `pairwise-mtbench`         | Pairwise   | `\boxed{A/B}`            | `parse_binary_boxed`    |
+| `pointwise-score`          | Pointwise  | `<score>N</score>`       | `parse_score`           |
+| `passthrough`              | Pointwise  | —                        | `None` (discriminative) |
+
+---
+
+## Config Reference
+
+All settings live in `judge_eval/hub/conf/hub_config.yaml` (generative) or `judge_eval/hub/conf/discriminative.yaml`. Every field can be overridden via Hydra CLI.
+
+### `run` — execution settings
+
+| Key           | Default   | Description                                             |
+|---------------|-----------|---------------------------------------------------------|
+| `seed`        | `42`      | Global RNG seed (affects shuffle_label, shuffle_sample) |
+| `save_dir`    | `results` | Root output directory                                   |
+| `log_level`   | `INFO`    | Logging level                                           |
+| `override`    | `false`   | If true, discard previous results and regenerate        |
+| `dry_run`     | `false`   | Print config summary without launching Ray              |
+| `show_prompt` | `false`   | Show formatted prompts (1 sample per dataset) and exit  |
+
+### `model` — model settings
+
+| Key                    | Default          | Description                                                                    |
+|------------------------|------------------|--------------------------------------------------------------------------------|
+| `path`                 | —                | HuggingFace model ID or local path                                             |
+| `backend`              | `generative`     | `generative` (vLLM generate) or `discriminative` (vLLM reward)                 |
+| `input_type`           | `pairwise`       | `pairwise` (compare A vs B) or `pointwise` (score each response independently) |
+| `dtype`                | `auto`           | Model dtype                                                                    |
+| `max_model_len`        | `null`           | Max context length; `null` = auto-detect                                       |
+| `custom_chat_template` | `null`           | Override the tokenizer's chat template                                         |
+| `enable_thinking`      | `false`          | Pass `enable_thinking` to `apply_chat_template`                                |
+| `template`             | `pairwise-fixed` | Prompt template name. Use `passthrough` for discriminative models              |
+| `parse_mode`           | `text`           | `text` (parse from generated text) or `logprob` (extract from logprobs)        |
+
+### `model.engine` — engine tuning
+
+| Key                      | Default | Description                         |
+|--------------------------|---------|-------------------------------------|
+| `gpu_memory_utilization` | `0.92`  | Fraction of GPU memory for KV cache |
+| `enable_prefix_caching`  | `true`  | Cache shared system-prompt KV       |
+| `enforce_eager`          | `false` | Disable CUDA graph capture          |
+
+### `model.generation` — sampling parameters (generative only)
+
+| Key           | Default | Description                      |
+|---------------|---------|----------------------------------|
+| `max_tokens`  | `8192`  | Max new tokens per response      |
+| `temperature` | `0.7`   | Sampling temperature             |
+| `top_p`       | `0.8`   | Nucleus sampling                 |
+| `top_k`       | `20`    | Top-k sampling                   |
+| `min_p`       | `0.0`   | Min-p sampling                   |
+| `n`           | `1`     | Number of generations per prompt |
+| `stop`        | `null`  | Stop sequences                   |
+
+### `resources` — pipeline resources
+
+| Key                 | Default         | Description                                                              |
+|---------------------|-----------------|--------------------------------------------------------------------------|
+| `gpu_groups`        | `[[0],[1],...]` | List of GPU-ID lists. One engine per sub-list. Sub-list length = TP size |
+| `engine_batch_size` | `512`           | Max requests per engine per inference call                               |
+| `n_workers`         | `8`             | CPU workers for misc tasks                                               |
+
+GPU group examples:
+
+```yaml
+gpu_groups: [[0],[1],[2],[3],[4],[5],[6],[7]]  # 8 × single-GPU
+gpu_groups: [[0,1,2,3],[4,5,6,7]]             # 2 × 4-GPU tensor-parallel
+gpu_groups: [[0,1],[2,3],[4,5],[6,7]]         # 4 × 2-GPU tensor-parallel
+```
+
+### `evaluation` — universal evaluation flags
+
+Defaults apply to all datasets. Per-dataset overrides are allowed.
+
+| Key               | Default | Description                                                |
+|-------------------|---------|------------------------------------------------------------|
+| `position_swap`   | `false` | Create swapped-position copies for consistency measurement |
+| `shuffle_label`   | `true`  | Randomly swap A/B assignment (pairwise only)               |
+| `shuffle_sample`  | `true`  | Shuffle sample order                                       |
+| `format_context`  | `true`  | Apply multi-turn `<user>/<assistant>` formatting to context |
+| `format_response` | `true`  | Wrap responses in `<assistant>` tags                       |
+| `max_samples`     | `null`  | Limit samples per dataset; `null` = all                    |
+
+### `datasets` — dataset list
+
+Simple string form or override dict form:
+
+```yaml
+datasets:
+  - nvidia/HelpSteer3              # simple: use all defaults
+  - allenai/reward-bench
+  - name: ScalerLab/JudgeBench     # override form
+    position_swap: true
+  - name: allenai/reward-bench-2
+    template: pairwise-fixed-concrete
+    max_samples: 500
+```
+
+Per-dataset override keys: `position_swap`, `shuffle_label`, `shuffle_sample`, `format_context`, `format_response`, `max_samples`, `template`.
+
+---
+
+## Run Tag & Resume
+
+### Run tag format
+
+```
+{template}_fc{format_context}_fr{format_response}_sl{shuffle_label}_seed{seed}_et{enable_thinking}
+```
+
+Example: `pairwise-fixed_fc1_fr1_sl1_seed42_etFalse`
+
+**Included** (changing any of these produces a new run):
+- `template`, `format_context`, `format_response`, `shuffle_label`, `seed`, `enable_thinking`
+
+**Excluded** (intentionally):
+- `position_swap` — excluded so original results are reused when enabling swap later, saving 50% of compute.
+- `parse_mode` — derived from the template, not independent.
+- Generation params (temperature, etc.) — use `run.override=true` to force regeneration if changed.
+
+### Resuming
+
+Every completed `UnitJudgment` is written to JSONL immediately. On re-run, completed units are skipped automatically.
+
+```bash
+judge-eval                   # resume — skips completed units
+judge-eval run.override=true # force fresh run
+```
+
+---
+
+## Output Structure
+
+```
+results/
+└── {model_name}/
+    └── {dataset_name}/
+        ├── {run_tag}.jsonl                     # per-unit checkpoint (resume)
+        ├── {run_tag}_metrics.json              # metric summary
+        └── {run_tag}_*_accuracy_table.txt      # per-dimension accuracy tables
+```
+
+### Aggregating results across models
+
+`generate_table.py` aggregates `_metrics.json` files across models and benchmarks:
+
+```bash
+python generate_table.py --format txt           # plain text
+python generate_table.py --format latex         # LaTeX
+python generate_table.py --format csv           # tab-separated (for spreadsheets)
+python generate_table.py --results_dir /path/to/results --output table.txt
+```
+
+Output columns: `Model | Overall Acc. | Avg Len | <sub-groups> | Pos. Cons.` (if `position_swap` was enabled).
+
+---
+
+## Extending
+
+### Adding a dataset
+
+1. **Create a YAML spec** in `judge_eval/data/dataset_specs/`:
+
+```yaml
+name: org/my-dataset           # HuggingFace dataset path
+split: test                    # HF split (null for all splits)
+adapter: my_dataset            # maps to judge_eval.data.adapters.extract_my_dataset
+unit_structure: pair           # pair | one_vs_many
+reduction: direct              # direct | all_correct | custom:module:fn
+aggregation:
+  overall: micro               # micro | macro:<dimension> | custom:module:fn
+  dimensions:
+    category: micro
+```
+
+2. **Write an adapter function** in `judge_eval/data/adapters.py`:
+
+```python
+def extract_my_dataset(example) -> EvalUnit:
+    return EvalUnit(
+        unit_id="",  # assigned by DatasetWorker
+        prompt=example["question"],
+        responses=[example["response_a"], example["response_b"]],
+        chosen_idx=0 if example["label"] == "a" else 1,
+        dimensions={"category": example["category"]},
+    )
+```
+
+3. **Add to config**: `datasets: [org/my-dataset]`
+
+#### Dataset spec fields
+
+| Field            | Required | Description                                                       |
+|------------------|----------|-------------------------------------------------------------------|
+| `name`           | Yes      | HuggingFace dataset path                                          |
+| `split`          | No       | HF split name (null = all splits)                                 |
+| `data_dir`       | No       | HF `data_dir` parameter                                           |
+| `adapter`        | Yes      | Name → `extract_{adapter}` in `judge_eval.data.adapters`          |
+| `unit_structure` | Yes      | `pair` (2 responses) or `one_vs_many` (1 chosen + N rejected)     |
+| `reduction`      | Yes      | `direct`, `all_correct`, or `custom:<module>:<fn>`                |
+| `aggregation`    | Yes      | Dict with `overall` and optional `dimensions`                     |
+| `filters`        | No       | List of filter expressions applied before the adapter             |
+
+#### Filters
+
+Applied to the raw HuggingFace dataset before the adapter. Form: `column op value`.
+
+```yaml
+filters:
+  - "preference_strength != 0"
+  - "split != train"
+  - "subset != Ties"
+```
+
+Supported operators: `==`, `!=`, `>`, `<`, `>=`, `<=`. Values are auto-parsed as int, float, or string.
+
+#### Reduction strategies
+
+- **`direct`**: Each pair independently graded. Pairwise: predict matches label. Pointwise: argmax of scores matches chosen.
+- **`all_correct`**: All comparisons for the unit must be correct (e.g., RewardBench-2: must beat all rejected responses).
+- **`custom:<module>:<fn>`**: Signature `(preferences, input_type, position_swap) -> UnitJudgment`.
+
+#### Aggregation strategies
+
+- **`micro`**: Simple average over all units.
+- **`macro:<dimension>`**: Average of per-group micro averages, grouped by the named dimension.
+- **`custom:<module>:<fn>`**: Signature `(judgments, aggregation_config, position_swap) -> dict`.
+
+---
+
+### Adding a template
+
+1. **Create a template class** in `judge_eval/templates/prompt_templates.py`:
+
+```python
+class my_custom_template(BasePromptTemplate):
+    def __init__(self):
+        system_msg = "You are a judge. Evaluate..."
+        user_msg = (
+            "Context: {context}\n"
+            "Response A: {response1}\n"   # pairwise: {response1}, {response2}
+            "Response B: {response2}\n"   # pointwise: {response}
+            "Verdict:"
+        )
+        super().__init__(system_msg, user_msg)
+```
+
+Standard field names:
+
+| Field        | Input Type | Description                       |
+|--------------|------------|-----------------------------------|
+| `{context}`  | Both       | The prompt / conversation context |
+| `{response1}`| Pairwise   | Response in the A position        |
+| `{response2}`| Pairwise   | Response in the B position        |
+| `{response}` | Pointwise  | Single response being evaluated   |
+
+Additional custom fields are populated from `EvalUnit.extra` (see below). The `passthrough` template bypasses the system/user prompt structure and builds raw `[user, assistant]` messages directly — use this for discriminative models.
+
+#### Injecting extra context into templates via `extra`
+
+Templates can reference any field from `EvalUnit.extra`. This lets adapters pass additional context (e.g., rubrics, feedback) into the prompt without modifying the pipeline. The key names in `extra` must match the `{field}` placeholders in the template exactly (e.g., `extra["rubric"]` → `{rubric}`).
+
+**Step 1: Adapter** — add fields to `extra`:
+
+```python
+def extract_my_dataset(example) -> EvalUnit:
+    extra = {
+        "rubric": example["rubric"],       # position-independent
+        "feedback": example["feedback"],   # position-dependent (references "Assistant A"/"Assistant B")
+        "_swap_markers": ["Assistant A", "Assistant B"],  # markers to swap when positions change
+    }
+    return EvalUnit(
+        unit_id="",
+        prompt=example["prompt"],
+        responses=[example["response_a"], example["response_b"]],
+        chosen_idx=0 if example["label"] == "A" else 1,
+        extra=extra,
+    )
+```
+
+**Step 2: Template** — reference `{rubric}` and `{feedback}` in the prompt:
+
+```python
+class pairwise_with_rubric_and_feedback(BasePromptTemplate):
+    def __init__(self):
+        system_msg = ""
+        user_msg = (
+            "[Context]\n{context}\n\n"
+            "[Assistant A]\n{response1}\n\n"
+            "[Assistant B]\n{response2}\n\n"
+            "Rubric:\n{rubric}\n\n"
+            "Feedback:\n{feedback}\n\n"
+            "Verdict (A or B) in <verdict>...</verdict>."
+        )
+        super().__init__(system_msg, user_msg)
+```
+
+**Position-aware marker swapping (`_swap_markers`):**
+
+When `shuffle_label` or `position_swap` swaps response positions, the expansion layer automatically swaps all marker occurrences in every string field of `extra`:
+- `rubric` has no markers → stays unchanged.
+- `feedback` contains "Assistant A" / "Assistant B" → references get swapped to match the new positions.
+
+Without `_swap_markers`, position-sensitive text would point to the wrong response after a swap. See `extract_helpsteer3_feedback` in `adapters.py` for a working example.
+
+2. **Register** in `judge_eval/templates/registry.py`:
+
+```python
+TEMPLATE_REGISTRY = {
+    ...
+    "my-custom": templates.my_custom_template,
+}
+```
+
+3. **Add a matching parser** (see below) and use in config: `model.template: my-custom`
+
+---
+
+### Adding a parser
+
+1. **Add a parser function** in `judge_eval/parsers/functions.py`:
+
+```python
+def parse_my_format(text: str, strict: bool = True) -> str | None:
+    """Returns "A", "B", or None on parse failure.
+    strict=True: only match structured tags.
+    strict=False: fall back to regex anywhere in text.
+    """
+    ...
+```
+
+2. **Register** in `judge_eval/parsers/registry.py` using the same key as the template:
+
+```python
+PARSER_REGISTRY = {
+    ...
+    "my-custom": parsers.parse_my_format,
+}
+```
+
+The parser key must match the template name — the parser is derived from the template automatically. For `passthrough` (discriminative models), no parser is needed.
+
+---
+
+## Future Work
+
+- **API backend** — support API-based judge/reward models (e.g., OpenAI, Anthropic) alongside vLLM.
+- **Custom dataset sources** — support local JSONL files and custom data loaders beyond HuggingFace.
+- **Majority voting** — `generation.n > 1`: generate N judgments per sample and take a majority vote.