Skip to content

eric-mitchell/direct-preference-optimization

View on GitHub ↗Related repositories →

Reference implementation for DPO (Direct Preference Optimization) on Eeyore

active 2024-11-012026-08-08 (UTC)

Partial coverage13,123 / 15,549 hourly files (84%) · 2 absent upstream · 2,420 failed, retryable2024-11-012026-08-11 (UTC)— sampled evenly across the window, so rankings and trends hold; absolute counts scale up.
Events
411
Pushes
0
Pull requests
2
Issues
2
Stars
359
Forks
39

Activity over time

Daily event counts in the loaded window

Line chart, 646 days from 2024-11-01 to 2026-08-08. Pushes: 0 total, peak 0 in a day. Pull requests: 2 total, peak 1 in a day. Issues: 2 total, peak 1 in a day. Comments: 9 total, peak 4 in a day. Stars: 359 total, peak 5 in a day.

  • Pushes
  • Pull requests
  • Issues
  • Comments
  • Stars

Stars, PRs, issues and forks are under-captured in the later part of this window. GH Archive progressively stopped capturing non-push events during 2026 — −95% or worse by the end of the window. Every series here except Pushes fades for that reason, so a decline above reflects the archive, not this repository. Pushes stay reliable throughout, so read them, and the contributor counts derived from them, as the real signal. Data health has the measurements.

Top contributors

Pushes, PRs, issues, reviews and comments — stars and forks excluded, so this is contribution rather than popularity

ContributorContributionsPushesPRsComments
Larry-u4012
yiyepiaoling07154004
Red-Scarff2001
HuXiangkun1001
xiaweijiexox1001
hknguyen201010

Recent activity

Latest issues, pull requests and releases

  • Issue comment#97Larry-u2025-09-25 03:02
    fix (preference_loss): dpo objective is inconsistent with original paper
  • Issue comment#98Larry-u2025-09-25 02:49
    [BUG] DPO objective is inconsistent with original paper
  • Issue#98Larry-u2025-09-25 02:49
    [BUG] DPO objective is inconsistent with original paper
  • Pull request#97Larry-u2025-09-25 01:55
  • Issue#96Red-Scarff2025-07-14 16:29
    When I use FSDPtrainer, I met a Bus error
  • Issue comment#42Red-Scarff2025-07-14 15:39
    llama7B issue
  • Pull request#95hknguyen202025-07-10 07:03
  • Issue comment#66xiaweijiexox2025-06-11 06:46
    Unable to Run SFT
  • Issue comment#68HuXiangkun2025-01-06 01:42
    Why does SFT sum the cross-entropy loss within each sequence?
  • Issue comment#39yiyepiaoling07152025-01-03 13:30
    Why sometimes chosen_rewards become negaive?
  • Issue comment#67yiyepiaoling07152025-01-03 13:12
    Using cross entropy loss to calculate DPO?
  • Issue comment#67yiyepiaoling07152025-01-03 13:11
    Using cross entropy loss to calculate DPO?
  • Issue comment#68yiyepiaoling07152025-01-03 13:08
    Why does SFT sum the cross-entropy loss within each sequence?

Totals cover only the window loaded into ClickHouse and count events, not GitHub's lifetime totals — 359 stars here means stars gained during the window, not the repo's star count.