[Data] Derive off-policy loss masks from vLLM render boundaries by WindChimeRan · Pull Request #794 · vllm-project/speculators

WindChimeRan · 2026-07-15T22:05:22Z

Purpose

Derive the off-policy training loss mask from a render boundary, with the vLLM instance as the single tokenization source. For each assistant turn we render the prompt and the prompt-plus-turn via /v1/chat/completions/render, and supervise the tokens the second render adds past the first.

Then we deprecate the fallback hf tokenization & fallback regex paths

This is the same prompt/completion boundary the engine already reports for on-policy regeneration (#729), reconstructed from the chat template. One tokenizer now produces the loss mask, the target hidden states, and the serving prompt — no separate local tokenization that can drift from serving.

Design write-up (worked examples, rationale): https://claude.ai/code/artifact/6cd3bf57-3d20-4a06-b27d-5b08e5fce6b4

What changes

Deleted: the two-tier local masking path — the HF {% generation %} tag-mask probe and the regex span-detector — plus the local render adapters and --assistant-pattern.
Added: a vLLM /render client and the boundary derivation.
Unchanged: on-policy pre-tokenized rows (Refactor(regen): emit generation boundary as loss mask for on-policy data, eliminate redundant re-tokenize roundtrip #729) still pass straight through.

Why token_ids, not the server mask: /render can also return assistant_tokens_mask, but it's only populated when the template carries {% generation %} tags, which the common models don't ship. We request only token_ids and compute the mask from the boundary — so it works on every template, and isn't blocked on the unreleased return_loss_mask (/render itself shipped in v0.15.0).

Requirement

--render-endpoint is now required for off-policy data (pre-tokenized input still works without it). Point it at the vLLM instance already used for hidden-state extraction, or a GPU-less vllm launch render. Single source of truth holds only if that instance's tokenizer_mode matches serving's.

How to review

~68% of the diff is pure deletion; the real surface is ~440 lines of source:

Bucket	What	How to read
Deletion (~1,560)	the old two-tier mask functions and the tests that covered them	skim
New (~530)	`render_client.py`, the boundary functions, `test_render_boundary.py`	review in isolation
Rewiring (~200)	`_preprocess_batch`, `build_eagle3_dataset`, `load_and_preprocess_dataset` switching to the boundary, plus the CLI swap	the logic change

The other preprocessing helpers are byte-identical to main. The delete-plus-rewire core is atomic; render_client.py is the one piece reviewable on its own.

Tests

# Serve any model with a chat template
vllm serve Qwen/Qwen3-0.6B --port 8000

# Derive masks from its render boundaries, on real data
python scripts/prepare_data.py \
  --model Qwen/Qwen3-0.6B \
  --data sharegpt \
  --render-endpoint http://localhost:8000 \
  --output /tmp/prepared \
  --max-samples 20

It prints the derived mask inline (blue = trainable, grey = masked):

<|im_start|>user
I have a location independent business that generates 1 miljon per year...<|im_end|>
<|im_start|>assistant
<think>

</think>

1/5: A clear target market and understanding of their needs is essential...<|im_end|>

The boundary is the thing to check: everything from <think> on is blue, while the user turn and the <|im_start|>assistant header stay grey.

Or automated — launches its own vLLM and asserts the masks:

pytest tests/e2e/smoke/test_render_boundary.py -v

Deferred / follow-ups

Concurrency — the boundary does 2–3 renders per turn, today riding datasets.map(num_proc). A batched/async client if throughput needs it.
Multimodal — prefix-stability not yet validated on an image model.
Reasoning emission — fan-out (on-policy, more rows) vs. a future single-row mode (cheaper, off-policy at turn seams); undecided.
Render-failure handling — a dead endpoint currently drains the dataset to empty per-row instead of aborting fast; should distinguish endpoint failure from a single bad conversation.
Shared tool-call parsing with regen ([Regen][tool call] Support tool calls in on-policy response regeneration #750), once both paths speak vLLM messages.

Checklist

Purpose of the PR
Test plan / results
Documentation update (CLI doc updated; user-guide tutorial pending)
I (a human) have reviewed this code to the best of my ability.

coderabbitai · 2026-07-15T22:05:30Z

Important

Review skipped

Auto incremental reviews are disabled on this repository.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro Plus

Run ID: 533bfb00-4d5f-4655-8678-88ccbdf304fe

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

🔍 Trigger review

📝 Walkthrough

Walkthrough

Changes

The preprocessing pipeline now uses a vLLM render endpoint to derive assistant loss boundaries, adds pre-tokenized passthrough handling, updates the CLI and public APIs, introduces an HTTP client, and replaces legacy assistant-pattern tests with render-boundary coverage.

Render-Based Dataset Preprocessing

Layer / File(s)	Summary
vLLM render client `pyproject.toml`, `src/speculators/data_generation/render_client.py`	Adds `httpx` and a retrying client for vLLM rendering requests, including response validation and HTTP error handling.
Render boundary construction `src/speculators/data_generation/preprocessing.py`	Adapts conversation content for vLLM, computes assistant-turn boundaries, and produces packed or fan-out loss-mask rows.
Preprocessing and CLI integration `src/speculators/data_generation/preprocessing.py`, `scripts/prepare_data.py`, `docs/cli/prepare_data.md`	Threads `render_endpoint` through preprocessing and dataset APIs, requires it for raw conversations, and documents the replacement CLI option. Pretokenized inputs bypass rendering.
Boundary and passthrough validation `tests/integration/datagen/test_render_boundary.py`, `tests/integration/datagen/test_preprocessing.py`, `tests/unit/scripts/test_response_regeneration.py`	Adds render-boundary, client-error, endpoint-validation, end-to-end, and pretokenized-path tests while removing obsolete assistant-pattern coverage.

Possibly related PRs

vllm-project/speculators#495: Both update multimodal conversation adaptation into vLLM-compatible message structures.
vllm-project/speculators#729: Both handle pre-tokenized samples through an _preprocess_batch passthrough path.
vllm-project/speculators#777: Both modify response-regeneration-related preprocessing tests and interfaces.

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Title check	✅ Passed	The title clearly summarizes the main change: deriving off-policy loss masks from vLLM render boundaries.
Description check	✅ Passed	The description is directly related to the changeset and accurately describes the render-boundary masking rewrite.

✨ Finishing Touches

🧪 Generate unit tests (beta)

Create PR with unit tests

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

_{Comment @coderabbitai help to get the list of available commands.}

coderabbitai

Actionable comments posted: 5

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@docs/cli/prepare_data.md`:
- Line 45: Update both off-policy conversation command examples in the
documentation to include the required --render-endpoint argument, using the same
vLLM server URL format described in the option reference. Apply the addition
consistently to the full example so all commands match the current CLI API.

In `@src/speculators/data_generation/preprocessing.py`:
- Around line 246-249: Update the preprocessing loop around the prompt_ids
length check to continue scanning after an over-length prompt instead of
breaking, so later history-rewritten contexts can still produce supervised
assistant rows when they fit within max_length. Add a regression test covering
an over-length turn followed by a shorter rewritten prompt and verify the later
turn is retained.
- Around line 415-418: Update the multimodal preprocessing path around
results["messages"] to preserve each row’s tools alongside messages. Extend the
data passed through vllm_client.ClientItem so the render path can send those
tools back when generating input_ids, preserving identical prompt-token
round-tripping for tool-use examples.

In `@src/speculators/data_generation/render_client.py`:
- Around line 50-56: Update the response classification around
InvalidResponseError so transient HTTP 429 responses remain eligible for
`@with_retries`, while deterministic client errors continue to short-circuit
retries. Exclude HTTPStatus.TOO_MANY_REQUESTS from the 4xx condition and
preserve the existing InvalidResponseError behavior for other covered statuses.
- Around line 62-67: Update the response validation around resp.json() and
token_ids to require a dictionary payload containing token_ids as a flat list of
plain non-negative integers; reject nested lists, floats, booleans, negatives,
and other invalid values with RenderError before returning. Preserve the
existing truncated-response context in the validation error and return only the
validated token_ids list.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

Push a commit to this branch (recommended)
Create a new PR with the fixes

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro

Run ID: b76f188b-c076-4959-845f-0db6b9380e9d

📥 Commits

Reviewing files that changed from the base of the PR and between 3251ed0 and 5a36d39.

📒 Files selected for processing (9)

docs/cli/prepare_data.md
pyproject.toml
scripts/prepare_data.py
src/speculators/data_generation/preprocessing.py
src/speculators/data_generation/render_client.py
tests/integration/datagen/test_preprocessing.py
tests/integration/datagen/test_regex_patterns.py
tests/integration/datagen/test_render_boundary.py
tests/unit/scripts/test_response_regeneration.py

💤 Files with no reviewable changes (2)

tests/integration/datagen/test_regex_patterns.py
tests/integration/datagen/test_preprocessing.py

Both examples pass `--data sharegpt`, which is off-policy, so as written they now fail immediately: ValueError: render_endpoint is required to derive loss masks for off-policy conversations. Pass --render-endpoint pointing at a vLLM server. The flag was documented in the argument reference but never reached the copy-pasteable commands. Note the requirement in Basic Usage as well, since it is conditional -- pre-tokenized input still needs no endpoint. Reported by CodeRabbit on vllm-project#794. Signed-off-by: Ranran Haoran Zhang <ranzhang@redhat.com>

408 and 429 sit inside the 4xx range, so they were classified as deterministic client errors and raised InvalidResponseError, which short-circuits @with_retries. A proxy shedding load therefore dropped the conversation permanently on the first 429 instead of backing off and retrying. Exclude them from the deterministic branch so they fall through to RenderError, which is retry-eligible. Everything else in 4xx keeps short-circuiting -- retrying a malformed request or a wrong URL wastes requests without changing the outcome. vLLM itself does not emit either status, but --render-endpoint accepts any URL and real deployments put proxies and load balancers in front of it. Tests: parametrized over 408/429, asserting the call is attempted max_retries+1 times. Both fail against the previous classification. Reported by CodeRabbit on vllm-project#794. Signed-off-by: Ranran Haoran Zhang <ranzhang@redhat.com>

408 and 429 fall inside 4xx, so they raised InvalidResponseError, which short-circuits @with_retries: a proxy shedding load dropped the conversation on the first 429 instead of backing off. Carve them out so they reach the retry-eligible RenderError; the rest of 4xx still short-circuits. vLLM emits neither, but --render-endpoint takes any URL and deployments front it with proxies that do. Reported by CodeRabbit on vllm-project#794. Signed-off-by: Ranran Haoran Zhang <ranzhang@redhat.com>

speculatorsbot

LGTM. Clean replacement of the two-tier local masking path (HF template probe + regex span detection) with a single render-boundary mechanism. The boundary derivation is correct, the LCP fallback for scaffold-divergent templates (DeepSeek-R1 / Qwen3.5) is load-bearing and well-validated, and the monotonicity argument for the early-exit break holds. Net deletion of ~1k lines. Agree with deferring the multimodal tools round-trip and the async/batched client to follow-ups. Recommend approving.

🤖 Generated with Claude Code using the /pr-review skill

mergify · 2026-07-17T16:30:43Z

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @WindChimeRan.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

Both examples pass `--data sharegpt`, which is off-policy, so as written they now fail immediately: ValueError: render_endpoint is required to derive loss masks for off-policy conversations. Pass --render-endpoint pointing at a vLLM server. The flag was documented in the argument reference but never reached the copy-pasteable commands. Note the requirement in Basic Usage as well, since it is conditional -- pre-tokenized input still needs no endpoint. Reported by CodeRabbit on vllm-project#794. Signed-off-by: Ranran Haoran Zhang <ranzhang@redhat.com>

408 and 429 fall inside 4xx, so they raised InvalidResponseError, which short-circuits @with_retries: a proxy shedding load dropped the conversation on the first 429 instead of backing off. Carve them out so they reach the retry-eligible RenderError; the rest of 4xx still short-circuits. vLLM emits neither, but --render-endpoint takes any URL and deployments front it with proxies that do. Reported by CodeRabbit on vllm-project#794. Signed-off-by: Ranran Haoran Zhang <ranzhang@redhat.com>

Replace the two-tier local masking path (HF {% generation %} assistant-mask probe + regex span detector) with a single mechanism: render each conversation through the training vLLM instance's /v1/chat/completions/render endpoint and take the loss mask from the boundary between the prompt render and the full render of each assistant turn. Append-only templates pack to one row per conversation; history-rewriting (reasoning) templates fan out to one row per turn. On-policy pre-tokenized rows (vllm-project#729) still pass straight through. One tokenizer now produces the mask, the target hidden states, and the serving prompt. Deletes ~270 lines of regex/tag masking and the local HF render path; adds render_client.py (token_ids only, no truncation) and boundary derivation. Signed-off-by: Ranran Haoran Zhang <ranzhang@redhat.com>

- drop redundant early-return in _render_boundary_rows (subsumed by the no-turns guard) - pass an is_multimodal bool into _preprocess_batch instead of the heavyweight processor, so the map closure no longer pickles it into every worker Signed-off-by: Ranran Haoran Zhang <ranzhang@redhat.com>

render_client.py imports httpx; it was only present transitively, so a clean install would fail to resolve it. Signed-off-by: Ranran Haoran Zhang <ranzhang@redhat.com>

Tests: - drop test_render_conversation_returns_token_ids (a one-line passthrough; the 200 path is covered by the missing-token_ids test) - drop test_context_filling_window_yields_no_rows (single-guard edge) - trim test_pretokenized_dataset_skips_render to its load-bearing assertion Prose: - condense the ported _render_boundary_rows docstring and the verbose comments/docstrings in preprocessing.py and render_client.py Signed-off-by: Ranran Haoran Zhang <ranzhang@redhat.com>

… vLLM Drop the packed branch. Every assistant turn now gets its own row carrying the history re-rendered the way inference sees it, for every template. Packing was an optimization, never a correctness requirement: it only held for append-only templates, and thinking models (Qwen3 strips <think> from history) already fanned out. Removing it deletes _Turn, the append_only chain check, and the two-pass structure -- one path instead of three. The cost lands on append-only templates, measured on 10 real multi-turn ShareGPT conversations rendered by Qwen2.5-0.5B-Instruct: packed 10 rows, 8606 tokens fan-out 38 rows, 20449 tokens (3.8x rows, 2.4x tokens) Qwen3-0.6B output is byte-identical before and after (37 rows), since it took the fan-out path already. The <think>-scaffold LCP fallback stays. It is load-bearing, not defensive: DeepSeek-R1-Distill pre-fills `<think>\n` in the generation prompt, and Qwen3.5 pre-fills an empty `<think></think>` that recorded reasoning then contradicts. Both break prefix-stability and land in that branch. Also warn when --seq-length silently truncates supervision. A row clipped mid-response keeps thousands of supervised tokens, so it passes the existing zero-supervision check and looks healthy while its tail and closing tokens are never learned. On 40 real gsm8k rows with recorded reasoning at --seq-length 2048, 13 are clipped, discarding a median 34% of the assistant turn (worst: 1482 of 3530 tokens) -- previously with no output at all. num_clipped counts only kept-but-truncated rows; rows clipped past their boundary already report as unsupervised and would otherwise be counted twice. Tests: - add tests/e2e/smoke/test_render_boundary.py -- the first test to exercise the render transport for real (URL, token_ids key, boundary from what vLLM actually returns). Qwen3-0.6B, live server, one conversation covering fan-out, the leading-assistant exclusion, and the trailing-user drop. - drop the tests that faked the render and are now covered there: the packed and fan-out routing tests, first-turn-context-only, trailing-dropped, and test_build_eagle3_dataset_packed_end_to_end (which named itself end-to-end while pointing at http://fake, and used a template that structurally could not reach fan-out). - keep what a live server cannot produce on demand: the scaffold fallback, the unstable guard, and the client's error paths. Signed-off-by: Ranran Haoran Zhang <ranzhang@redhat.com>

Both examples pass `--data sharegpt`, which is off-policy, so as written they now fail immediately: ValueError: render_endpoint is required to derive loss masks for off-policy conversations. Pass --render-endpoint pointing at a vLLM server. The flag was documented in the argument reference but never reached the copy-pasteable commands. Note the requirement in Basic Usage as well, since it is conditional -- pre-tokenized input still needs no endpoint. Reported by CodeRabbit on vllm-project#794. Signed-off-by: Ranran Haoran Zhang <ranzhang@redhat.com>

- `assert sum(mask) > 0`: _append_row returns "unsupervised" and does not append when num_valid_tokens == 0, so every row reaching the dataset satisfies this by construction. - `assert len(ids) == len(mask)`: the mask is built as [0]*boundary + [1]*(len(full_ids) - boundary) and both are clipped by the same slice, so they cannot diverge -- and `zip(..., strict=True)` two lines below already raises on skew. - unused `logging` import and module logger. Signed-off-by: Ranran Haoran Zhang <ranzhang@redhat.com>

408 and 429 fall inside 4xx, so they raised InvalidResponseError, which short-circuits @with_retries: a proxy shedding load dropped the conversation on the first 429 instead of backing off. Carve them out so they reach the retry-eligible RenderError; the rest of 4xx still short-circuits. vLLM emits neither, but --render-endpoint takes any URL and deployments front it with proxies that do. Reported by CodeRabbit on vllm-project#794. Signed-off-by: Ranran Haoran Zhang <ranzhang@redhat.com>

speculatorsbot

Re-reviewed after the 7 commits since my previous pass (ab44652..63cfcf5). The always-fan-out simplification is a good call -- removing the packed/append-only branch deletes _Turn, the two-pass chain check, and the dual routing logic in exchange for a measured 2.4x token cost on append-only templates, which is acceptable given the complexity reduction. The e2e test against a live Qwen3-0.6B server (test_render_boundary.py) is the most valuable addition: it is the only test that exercises the actual /render transport and catches contract drift, which the stubbed unit tests structurally cannot. The transient-4xx retry fix (TRANSIENT_STATUSES for 408/429) correctly threads through the existing with_retries/InvalidResponseError short-circuit. All coderabbit findings from the first round are either addressed in these commits or explicitly deferred (multimodal tools round-trip). DCO sign-offs present on all 8 commits. LGTM. Recommend approving.

🤖 Generated with Claude Code using the /pr-review skill

orestis-z

That's a great delta -1.1k lines :)

Some nits on render_client.py (feel free to ignore them)

orestis-z

Another round of style nits (feel free to ignore and/or pick some to fix you agree with / add value)

orestis-z · 2026-07-21T13:38:10Z

+    num_convs_in = 0
+    num_convs_empty = 0
+
    for idx, conv in enumerate(conversations):


nit: this method seems to be doing quite a lot, and there's a loop inside a loop. Could extract helper method(s) to hide some of the logic.

Not sure how to do that cleanly.

…ports Signed-off-by: Ranran Haoran Zhang <ranzhang@redhat.com>

Signed-off-by: Ranran Haoran Zhang <ranzhang@redhat.com>

…sation Signed-off-by: Ranran Haoran Zhang <ranzhang@redhat.com>

Signed-off-by: Ranran Haoran Zhang <ranzhang@redhat.com>

mergify · 2026-07-24T03:31:48Z

Merge Protections

🔴 1 of 1 protections blocking · waiting on 👀 reviews

	Protection	Waiting on
🔴	Require approval from approved reviewers list	👀 reviews

🔴 Require approval from approved reviewers list

Waiting for any of

approved-reviews-by = dsikka
approved-reviews-by = fynnsu
approved-reviews-by = orestis-z
approved-reviews-by = rahul-tuli
approved-reviews-by = shanjiaz

This rule is failing.

All pull requests must have at least one approving review from a member of the approved reviewers list before merging.

shanjiaz

Taking a look now. Sorry about the delay.

…ing_dataset The prepared dataset is not eagle3-specific: prepare_data.py takes no speculator type, and train.py builds eagle3, dflash, dspark, peagle, and mtp from the same rows. Rename the function and drop EAGLE3 from the two docstrings that described its output. Signed-off-by: Ranran Haoran Zhang <ranzhang@redhat.com>

Point --render-endpoint at the launch_vllm.py instance rather than a second server, and flag the URL-form difference: data_generation_offline takes an /v1-suffixed endpoint, while this one is appended to and 404s on that form. Signed-off-by: Ranran Haoran Zhang <ranzhang@redhat.com>

The argparse help still carried the pre-change wording. Also scope the requirement correctly: --data is repeatable and the check runs per dataset, so a single raw input among pre-tokenized ones still needs the endpoint. Signed-off-by: Ranran Haoran Zhang <ranzhang@redhat.com>

mergify · 2026-07-26T15:54:06Z

The quality checks have failed. Please run make style and make quality under
the root directory to address the lint failures. You will need to install the
dev optional install to get the required linting packages:
https://github.com/vllm-project/speculators/blob/main/CONTRIBUTING.md

shanjiaz

I think the diff makes sense and checking the diff for loss masks is indeed smart. Verified locally. Could you please address Orestis' style concerns? we could merge soon. Would want to make sure this works on multi-modal data tho. Especially the coco dataset we claim to support.

mypy rejects passing None where ProcessorLike is expected. Neither test reads the processor: the missing-endpoint guard raises before it is used, and pre-tokenized rows skip preprocessing. Name that with a cast sentinel rather than an ignore. Signed-off-by: Ranran Haoran Zhang <ranzhang@redhat.com>

mergify Bot added the documentation Improvements or additions to documentation label Jul 15, 2026

WindChimeRan force-pushed the feat/render-boundary-loss-mask branch 3 times, most recently from e5f7fb8 to 5a36d39 Compare July 16, 2026 05:26

WindChimeRan marked this pull request as ready for review July 16, 2026 06:15

coderabbitai Bot reviewed Jul 16, 2026

View reviewed changes

WindChimeRan marked this pull request as draft July 16, 2026 08:05

WindChimeRan force-pushed the feat/render-boundary-loss-mask branch from d29cad4 to 7ed84a4 Compare July 17, 2026 03:26

WindChimeRan mentioned this pull request Jul 17, 2026

Apply max_length truncation on HF assistant-mask path #645

Closed

4 tasks

WindChimeRan marked this pull request as ready for review July 17, 2026 04:07

WindChimeRan force-pushed the feat/render-boundary-loss-mask branch from 19bcfc7 to ab44652 Compare July 17, 2026 04:21

WindChimeRan mentioned this pull request Jul 17, 2026

Add --render-endpoint to prepare_data.py for vLLM-delegated tokenization (#652) #665

Closed

4 tasks

speculatorsbot reviewed Jul 17, 2026

View reviewed changes

mergify Bot added the needs-rebase label Jul 17, 2026

WindChimeRan force-pushed the feat/render-boundary-loss-mask branch from ab44652 to c0396ee Compare July 17, 2026 17:52

WindChimeRan added 8 commits July 17, 2026 12:54

build(deps): declare httpx (used by render_client)

edba779

render_client.py imports httpx; it was only present transitively, so a clean install would fail to resolve it. Signed-off-by: Ranran Haoran Zhang <ranzhang@redhat.com>

WindChimeRan force-pushed the feat/render-boundary-loss-mask branch from c0396ee to 63cfcf5 Compare July 17, 2026 17:55

mergify Bot removed the needs-rebase label Jul 17, 2026

speculatorsbot reviewed Jul 21, 2026

View reviewed changes

orestis-z self-requested a review July 21, 2026 11:53

orestis-z reviewed Jul 21, 2026

View reviewed changes

Comment thread src/speculators/data_generation/render_client.py Outdated

Comment thread src/speculators/data_generation/render_client.py

Comment thread src/speculators/data_generation/render_client.py Outdated

Comment thread tests/integration/datagen/test_render_boundary.py Outdated

orestis-z reviewed Jul 21, 2026

View reviewed changes

WindChimeRan added 4 commits July 22, 2026 00:22

test(data): spell out the module names in the render-boundary test im…

6fa3aad

…ports Signed-off-by: Ranran Haoran Zhang <ranzhang@redhat.com>

style(data): initialize each counter on its own line

8edbe76

Signed-off-by: Ranran Haoran Zhang <ranzhang@redhat.com>

style(data): drop the redundant body dict annotation

7d39813

Signed-off-by: Ranran Haoran Zhang <ranzhang@redhat.com>

fix(data): report render endpoint errors in full instead of truncating

3e0ba1c

Signed-off-by: Ranran Haoran Zhang <ranzhang@redhat.com>

WindChimeRan force-pushed the feat/render-boundary-loss-mask branch from fc903f3 to 3e0ba1c Compare July 22, 2026 06:10

WindChimeRan added 3 commits July 22, 2026 01:44

fix(data): skip only the over-length turn, not the rest of the conver…

96ee32d

…sation Signed-off-by: Ranran Haoran Zhang <ranzhang@redhat.com>

refactor(data): type the boundary rows and trim comments

c3fb82f

Signed-off-by: Ranran Haoran Zhang <ranzhang@redhat.com>

fix(data): annotate conversations as a batch of conversations

11a1fc1

Signed-off-by: Ranran Haoran Zhang <ranzhang@redhat.com>

This comment was marked as resolved.

Sign in to view

shanjiaz reviewed Jul 26, 2026

View reviewed changes

Comment thread docs/cli/prepare_data.md Outdated

shanjiaz reviewed Jul 26, 2026

View reviewed changes

Comment thread src/speculators/data_generation/preprocessing.py Outdated

WindChimeRan requested review from dsikka, fynnsu and rahul-tuli as code owners July 26, 2026 06:28

WindChimeRan added 2 commits July 26, 2026 01:30

mergify Bot added the quality-failed label Jul 26, 2026

shanjiaz reviewed Jul 26, 2026

View reviewed changes

Uh oh!

Conversation

WindChimeRan commented Jul 15, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Purpose

What changes

Requirement

How to review

Tests

Deferred / follow-ups

Checklist

Uh oh!

coderabbitai Bot commented Jul 15, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Review skipped

Walkthrough

Changes

Uh oh!

coderabbitai Bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

speculatorsbot left a comment

Choose a reason for hiding this comment

Uh oh!

mergify Bot commented Jul 17, 2026

Uh oh!

speculatorsbot left a comment

Choose a reason for hiding this comment

Uh oh!

orestis-z left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

orestis-z left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

orestis-z Jul 21, 2026

Choose a reason for hiding this comment

Uh oh!

WindChimeRan Jul 24, 2026

Choose a reason for hiding this comment

Uh oh!

This comment was marked as resolved.

Uh oh!

mergify Bot commented Jul 24, 2026

Merge Protections

🔴 Require approval from approved reviewers list

Uh oh!

Uh oh!

shanjiaz left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

mergify Bot commented Jul 26, 2026

Uh oh!

shanjiaz left a comment

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

WindChimeRan commented Jul 15, 2026 •

edited

Loading

coderabbitai Bot commented Jul 15, 2026 •

edited

Loading

orestis-z left a comment •

edited

Loading

shanjiaz left a comment •

edited

Loading