Repository navigation
Add single-node Qwen3 DAPO math example - #3951
Merged
felipemello1 merged 10 commits intoJul 21, 2026
Merged
Conversation
felipemello1
requested review from
fegin,
tianyu-l,
wconstab and
wwwjn
as code owners
July 20, 2026 20:18
meta-cla
Bot
added
the
CLA Signed
felipemello1
force-pushed
the
feature/qwen3-4b-dapo-8k
branch
from
July 21, 2026 01:22
78cab15 to
4b14a75
Compare
felipemello1
force-pushed
the
feature/qwen3-4b-dapo-8k
branch
from
July 21, 2026 03:00
4b14a75 to
bb8e1d3
Compare
tianyu-l
reviewed
Jul 21, 2026
Comment on lines
+21
to
+22
| # Rule-based verification for math RL examples | ||
| "math-verify==0.9.0", |
Contributor
There was a problem hiding this comment.
This should be recipe / example level of requirements, rather than repo-wise? Some one who works on pre-training / diffusion models may not need this lib.
If you need to put this in CI, we can do something similar to https://github.com/pytorch/torchtitan/blob/main/.ci/docker/requirements-vlm.txt
Contributor
There was a problem hiding this comment.
Do we need this? How should I appreciate this vs. https://github.com/felipemello1/torchtitan/blob/main/torchtitan/experiments/rl/train.py
tianyu-l
reviewed
Jul 21, 2026
Comment on lines
+58
to
+67
| ## 150-step result | ||
|
|
||
| The single-node TP=2 trial completed optimizer steps 0 through 149 in 1 hour 46 minutes. This trial used the earlier 7,168-token response and 9,216-token packing limits; the runnable configuration above restores the full 8,192-token response budget. | ||
|
|
||
| ```text | ||
| metric first 10 steps last 10 steps | ||
| rollout reward, mean 0.067 0.401 | ||
| rollout total length, mean 988 tokens 3,284 tokens | ||
| rollout truncation rate, mean 1.75% 14.55% | ||
| ``` |
Contributor
There was a problem hiding this comment.
Could you share a wandb log on the real job?
added 6 commits
July 21, 2026 10:45
Summary: Test Plan: Reviewers: Subscribers: Tasks: Tags:
tianyu-l
reviewed
Jul 21, 2026
Comment on lines
+96
to
+98
|  | ||
|
|
||
|  |
Contributor
There was a problem hiding this comment.
Curves in these pictures would be easy to get outdated and not reproducible on later commit. Maybe only add them in PR summary, or add a commit pin to the results.
tianyu-l
approved these changes
Jul 21, 2026
yichuan-w
reviewed
Jul 21, 2026
|
|
||
| Training uses the 12,643-row [filtered DAPO-Math dataset](https://huggingface.co/datasets/hamishivi/DAPO-Math-17k-Processed_filtered). Each row contains one user prompt and its verifiable final answer. | ||
|
|
||
| Validation uses all 30 problems from [AIME 2025](https://huggingface.co/datasets/opencompass/AIME2025). The same single-turn environment and Math-Verify reward are used for training and validation. |
Member
There was a problem hiding this comment.
can we also put a validation set accuracy figure in the readme so people can be more convinced
Contributor
Author
There was a problem hiding this comment.
yes, i ahve to rerun it. I left as a todo at the bottom of the readme
Member
|
LGTM, very clean PR |
yichuan-w
pushed a commit
to yichuan-w/torchtitan
that referenced
this pull request
Jul 28, 2026
## Summary - Add a single-node Qwen3-4B-Base DAPO-Math example with a TP=2 trainer and six independent TP=1 generators. - Train on the 12,643-row filtered DAPO-Math dataset and validate on AIME 2025 with Math-Verify rewards. - Document the recipe and the measured 150-step training result. - Smoke run: https://meta.wandb.io/felipemello/titan_rl/runs/3xhecewt <img width="524" height="313" alt="image" src="https://github.com/user-attachments/assets/b45f68a1-f3f4-4a1f-be45-f0c4ad6515f7" /> <img width="517" height="315" alt="image" src="https://github.com/user-attachments/assets/252b3da4-30d8-4b5d-b4df-ec518efb9385" /> --------- Co-authored-by: Felipe Mello <felipemello@fb.com>
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
3 participants
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