Open
Conversation
WalkthroughAdds two documentation resources: an OSFT end-to-end tutorial notebook with dataset creation, model/config examples, distributed training setup, run wrapper, and post-training inspection; and a training_hub fine-tuning guide covering installation, JSONL data format, SFT/OSFT loss-masking behavior, parameter reference, and multi-node launch notes. Changes
Sequence Diagram(s)(omitted — changes are documentation-only and do not introduce new multi-component control flow) Estimated code review effort🎯 2 (Simple) | ⏱️ ~12 minutes Suggested reviewers
Poem
🚥 Pre-merge checks | ✅ 3✅ Passed checks (3 passed)
✏️ Tip: You can configure your own custom pre-merge checks in the settings. ✨ Finishing Touches📝 Generate docstrings
🧪 Generate unit tests (beta)
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. Comment |
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 7
🤖 Prompt for all review comments with AI agents
Verify each finding against the current code and only fix it if needed.
Inline comments:
In `@docs/en/workbench/how_to/osft_comprehensive_tutorial.ipynb`:
- Line 402: Remove the internal author TODO string "# TODO: revert these
overrides after we've concluded training" from the notebook cell in the tutorial
flow; either delete that line or replace it with user-facing instructions that
explain the current state (e.g., "Temporary overrides applied for training — no
action required") so the published tutorial contains only user-directed
guidance; locate the cell containing that exact TODO string to make the edit
(search for "# TODO: revert these overrides after we've concluded training").
- Line 886: The docs show conflicting defaults for the setting 'use_liger'—the
table row lists default False while the configuration cell assigns use_liger =
True and the companion MDX page also differs; pick the canonical default (e.g.,
True) and update all occurrences to match: change the table/default example
entry to the chosen value and/or update the configuration cell to the chosen
value so the table row, the config cell (variable use_liger), and the companion
MDX description all state the same default and example.
- Around line 539-552: The multi-node template entries (keys like
"multi_node_master" and "multi_node_worker") are inconsistent: change "nnodes"
to match the described 4-node topology (set nnodes: 4 for master and workers),
update "node_rank" comments so workers indicate ranks 1..3, and make
"rdzv_endpoint" a clearly labeled placeholder (e.g., "MASTER_IP:PORT") rather
than a concrete cluster IP; keep "rdzv_id" unchanged but ensure descriptions
(the "description" fields) reflect 4 nodes and indicate that workers must change
their node_rank per node.
- Around line 297-299: The preset is inconsistent: the JSON fields model_name
and model_path disagree (model_name currently "Qwen 2.5 7B Instruct" while
model_path points to "Qwen3-0.6B"); update one to match the other (e.g., set
model_name to "Qwen3-0.6B" if model_path is correct, or change model_path to the
correct HuggingFace/local path for "Qwen 2.5 7B Instruct"), and ensure the
adjacent comment and example_unfreeze_rank_ratio remain accurate for the chosen
model.
In `@docs/en/workbench/how_to/training_hub_fine_tuning.mdx`:
- Line 136: The max_seq_len value in the OSFT example is a typo: update the
assignment of max_seq_len from 8196 to 8192 in the OSFT/example config (look for
the max_seq_len=8196 occurrence) so it matches the 8K examples elsewhere and
maintain consistency across docs.
- Line 23: The table row under the "Catastrophic forgetting" column uses awkward
modal phrasing "Requires mixed/replay data to mitigate"; change it to a clearer
verb form such as "Mitigated by using mixed/replay data" (or "Mitigated using
mixed/replay data") so the row reads: | **Catastrophic forgetting** | Mitigated
by using mixed/replay data | Prevented algorithmically |, updating the cell text
accordingly.
ℹ️ Review info
⚙️ Run configuration
Configuration used: Organization UI
Review profile: CHILL
Plan: Pro
Run ID: 380e0c93-7883-4a94-b58a-fde7a5e43c12
📒 Files selected for processing (3)
docs/en/workbench/how_to/osft_comprehensive_tutorial.ipynbdocs/en/workbench/how_to/sft_comprehensive_tutorial.ipynbdocs/en/workbench/how_to/training_hub_fine_tuning.mdx
There was a problem hiding this comment.
♻️ Duplicate comments (1)
docs/en/workbench/how_to/training_hub_fine_tuning.mdx (1)
23-23:⚠️ Potential issue | 🟡 MinorUse clearer phrasing for catastrophic forgetting mitigation.
Line 23 reads awkwardly (“Requires mixed/replay data to mitigate”). Please rephrase for clarity, e.g. “Mitigated by using mixed/replay data.”
Suggested edit
-| **Catastrophic forgetting** | Requires mixed/replay data to mitigate | Prevented algorithmically | +| **Catastrophic forgetting** | Mitigated by using mixed/replay data | Prevented algorithmically |🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@docs/en/workbench/how_to/training_hub_fine_tuning.mdx` at line 23, Update the table row in the file where the "Catastrophic forgetting" entry appears (the cell currently reads "Requires mixed/replay data to mitigate") to a clearer phrasing such as "Mitigated by using mixed/replay data" so the mitigation column reads concisely and reads naturally; locate the exact string "Catastrophic forgetting" in the table and replace the adjacent mitigation cell accordingly.
🤖 Prompt for all review comments with AI agents
Verify each finding against the current code and only fix it if needed.
Duplicate comments:
In `@docs/en/workbench/how_to/training_hub_fine_tuning.mdx`:
- Line 23: Update the table row in the file where the "Catastrophic forgetting"
entry appears (the cell currently reads "Requires mixed/replay data to
mitigate") to a clearer phrasing such as "Mitigated by using mixed/replay data"
so the mitigation column reads concisely and reads naturally; locate the exact
string "Catastrophic forgetting" in the table and replace the adjacent
mitigation cell accordingly.
ℹ️ Review info
⚙️ Run configuration
Configuration used: Organization UI
Review profile: CHILL
Plan: Pro
Run ID: 6df118c7-39ae-4e8a-8818-15adca701633
📒 Files selected for processing (2)
docs/en/workbench/how_to/osft_comprehensive_tutorial.ipynbdocs/en/workbench/how_to/training_hub_fine_tuning.mdx
✅ Files skipped from review due to trivial changes (1)
- docs/en/workbench/how_to/osft_comprehensive_tutorial.ipynb
There was a problem hiding this comment.
Actionable comments posted: 1
♻️ Duplicate comments (1)
docs/en/workbench/how_to/training_hub_fine_tuning.mdx (1)
23-23:⚠️ Potential issue | 🟡 MinorFix awkward modal phrasing in the table row.
The phrasing "Requires mixed/replay data to mitigate" is grammatically awkward. Consider "Mitigated using mixed/replay data" or "Needs mixed/replay data to mitigate."
📝 Suggested edit
-| **Catastrophic forgetting** | Requires mixed/replay data to mitigate | Prevented algorithmically | +| **Catastrophic forgetting** | Mitigated using mixed/replay data | Prevented algorithmically |🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@docs/en/workbench/how_to/training_hub_fine_tuning.mdx` at line 23, Update the table row that currently reads "Requires mixed/replay data to mitigate" to a clearer phrasing such as "Mitigated using mixed/replay data" (or alternatively "Needs mixed/replay data to mitigate") in the docs/en/workbench/how_to/training_hub_fine_tuning.mdx table so the modal phrasing is grammatically correct and consistent with other rows; replace the exact cell text in the row labelled "Catastrophic forgetting" with the chosen suggestion.
🧹 Nitpick comments (2)
docs/en/workbench/how_to/training_hub_fine_tuning.mdx (2)
45-51: Clarify thepretrainingrole.Line 45 lists
pretrainingas a supported role value, but the masking behavior section (lines 47-51) doesn't explain what this role does or how it differs from using the"unmask": trueflag. Consider adding a brief explanation.📝 Suggested addition
After line 45, add:
- `pretraining` role: Content with this role is always included in the loss calculation, similar to enabling `unmask` for the sample.Or integrate into the masking behavior section.
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@docs/en/workbench/how_to/training_hub_fine_tuning.mdx` around lines 45 - 51, Add a brief explanation of the "pretraining" role and how it differs from the other roles: state that the `pretraining` role means content marked with role="pretraining" is always included in the loss (equivalent to setting `"unmask": true` for that sample), and mention how it relates to SFT and OSFT masking (i.e., SFT masks assistant-only by default, OSFT uses `unmask_messages`, while `pretraining` forces inclusion regardless of those flags). Reference the `pretraining` role string and the `"unmask": true` flag and `unmask_messages` parameter so readers can find the relevant behavior descriptions.
98-99: Complete the guidance forunfreeze_rank_ratiorange.The comment provides guidance for 0.1–0.5 but doesn't explain what values above 0.5 represent, leaving users uncertain about the upper half of the valid range (0.5–1.0).
📝 Suggested enhancement
-unfreeze_rank_ratio = 0.25 # 0.1–0.3 conservative, 0.3–0.5 balanced +unfreeze_rank_ratio = 0.25 # 0.1–0.3 conservative, 0.3–0.5 balanced, 0.5–1.0 aggressive🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@docs/en/workbench/how_to/training_hub_fine_tuning.mdx` around lines 98 - 99, The documentation for unfreeze_rank_ratio stops at 0.5 and leaves the upper half of the valid range undefined; update the guidance for the unfreeze_rank_ratio parameter (used in the training/fine-tuning docs) to explain what values in 0.5–1.0 mean for training behavior and trade-offs (e.g., >0.5 progressively unfreezes more rank components, increasing model plasticity and compute/memory requirements, higher risk of overfitting but faster adaptation), and give concrete recommendations for when to choose conservative (0.1–0.3), balanced (0.3–0.5), and aggressive (0.5–1.0) settings along with a short sentence on pitfalls (compute, memory, overfitting) to help users pick a value.
🤖 Prompt for all review comments with AI agents
Verify each finding against the current code and only fix it if needed.
Inline comments:
In `@docs/en/workbench/how_to/training_hub_fine_tuning.mdx`:
- Around line 60-61: The two table links for "SFT Comprehensive Tutorial" and
"OSFT Comprehensive Tutorial" currently point to the directory view; update
their URLs so they reference the actual notebook files
(sft_comprehensive_tutorial.ipynb and osft_comprehensive_tutorial.ipynb) using
either the /blob/master/ path for direct file view or /raw/master/ for direct
download, ensuring the markdown link targets the specific file rather than the
folder.
---
Duplicate comments:
In `@docs/en/workbench/how_to/training_hub_fine_tuning.mdx`:
- Line 23: Update the table row that currently reads "Requires mixed/replay data
to mitigate" to a clearer phrasing such as "Mitigated using mixed/replay data"
(or alternatively "Needs mixed/replay data to mitigate") in the
docs/en/workbench/how_to/training_hub_fine_tuning.mdx table so the modal
phrasing is grammatically correct and consistent with other rows; replace the
exact cell text in the row labelled "Catastrophic forgetting" with the chosen
suggestion.
---
Nitpick comments:
In `@docs/en/workbench/how_to/training_hub_fine_tuning.mdx`:
- Around line 45-51: Add a brief explanation of the "pretraining" role and how
it differs from the other roles: state that the `pretraining` role means content
marked with role="pretraining" is always included in the loss (equivalent to
setting `"unmask": true` for that sample), and mention how it relates to SFT and
OSFT masking (i.e., SFT masks assistant-only by default, OSFT uses
`unmask_messages`, while `pretraining` forces inclusion regardless of those
flags). Reference the `pretraining` role string and the `"unmask": true` flag
and `unmask_messages` parameter so readers can find the relevant behavior
descriptions.
- Around line 98-99: The documentation for unfreeze_rank_ratio stops at 0.5 and
leaves the upper half of the valid range undefined; update the guidance for the
unfreeze_rank_ratio parameter (used in the training/fine-tuning docs) to explain
what values in 0.5–1.0 mean for training behavior and trade-offs (e.g., >0.5
progressively unfreezes more rank components, increasing model plasticity and
compute/memory requirements, higher risk of overfitting but faster adaptation),
and give concrete recommendations for when to choose conservative (0.1–0.3),
balanced (0.3–0.5), and aggressive (0.5–1.0) settings along with a short
sentence on pitfalls (compute, memory, overfitting) to help users pick a value.
ℹ️ Review info
⚙️ Run configuration
Configuration used: Organization UI
Review profile: CHILL
Plan: Pro
Run ID: c7e52585-b4c7-4a3c-abed-718a1e780ed0
📒 Files selected for processing (1)
docs/en/workbench/how_to/training_hub_fine_tuning.mdx
Comment on lines
+60
to
+61
| | SFT Comprehensive Tutorial | Supervised Fine-Tuning | [Download sft_comprehensive_tutorial.ipynb](https://github.com/alauda/aml-docs/tree/master/docs/en/workbench/how_to) | | ||
| | OSFT Comprehensive Tutorial | Orthogonal Subspace Fine-Tuning | [Download osft_comprehensive_tutorial.ipynb](https://github.com/alauda/aml-docs/tree/master/docs/en/workbench/how_to) | |
There was a problem hiding this comment.
Fix incomplete download links.
Both download links point to the same directory view (/tree/master/docs/en/workbench/how_to) rather than to specific notebook files. Users expecting to download sft_comprehensive_tutorial.ipynb and osft_comprehensive_tutorial.ipynb will land on a directory listing instead.
🔗 Suggested fix
Update the links to point to the actual files:
-| SFT Comprehensive Tutorial | Supervised Fine-Tuning | [Download sft_comprehensive_tutorial.ipynb](https://github.com/alauda/aml-docs/tree/master/docs/en/workbench/how_to) |
-| OSFT Comprehensive Tutorial | Orthogonal Subspace Fine-Tuning | [Download osft_comprehensive_tutorial.ipynb](https://github.com/alauda/aml-docs/tree/master/docs/en/workbench/how_to) |
+| SFT Comprehensive Tutorial | Supervised Fine-Tuning | [Download sft_comprehensive_tutorial.ipynb](https://github.com/alauda/aml-docs/tree/master/docs/en/workbench/how_to/sft_comprehensive_tutorial.ipynb) |
+| OSFT Comprehensive Tutorial | Orthogonal Subspace Fine-Tuning | [Download osft_comprehensive_tutorial.ipynb](https://github.com/alauda/aml-docs/tree/master/docs/en/workbench/how_to/osft_comprehensive_tutorial.ipynb) |Or use /blob/master/ for direct file view, or /raw/master/ for raw download.
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed.
In `@docs/en/workbench/how_to/training_hub_fine_tuning.mdx` around lines 60 - 61,
The two table links for "SFT Comprehensive Tutorial" and "OSFT Comprehensive
Tutorial" currently point to the directory view; update their URLs so they
reference the actual notebook files (sft_comprehensive_tutorial.ipynb and
osft_comprehensive_tutorial.ipynb) using either the /blob/master/ path for
direct file view or /raw/master/ for direct download, ensuring the markdown link
targets the specific file rather than the folder.
Deploying alauda-ai with
|
| Latest commit: |
ce9309d
|
| Status: | ✅ Deploy successful! |
| Preview URL: | https://07f2b7a9.alauda-ai.pages.dev |
| Branch Preview URL: | https://add-docs-for-training-hub.alauda-ai.pages.dev |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary by CodeRabbit