pith. machine review for the scientific record. sign in
Pith Number

pith:NUOAWJTF

pith:2024:NUOAWJTF3Y3ZFVGJIAJROENBKH
not attested not anchored not stored refs resolved

Smaug: Fixing Failure Modes of Preference Optimisation with DPO-Positive

Arka Pal, Colin White, Deep Karkhanis, Manley Roberts, Samuel Dooley, Siddartha Naidu

Standard DPO can lower the absolute likelihood of preferred responses while still raising their relative odds; a modified loss called DPOP prevents the drop and improves results.

arxiv:2402.13228 v2 · 2024-02-20 · cs.CL · cs.AI · cs.LG

Record completeness

1 Bitcoin timestamp
2 Internet Archive
3 Author claim open · sign in to claim
4 Citations open
5 Replications open

Claims

C1strongest claim

we find that DPOP outperforms DPO and other fine-tuning procedures across a wide variety of datasets and downstream tasks, including datasets with high edit distances between completions. Furthermore, we find that the DPOP-tuned model outperforms the DPO-tuned model (all else equal) on benchmarks independent of the fine-tuning data, such as MT-Bench. Finally, using DPOP, we create and open-source Smaug-34B and Smaug-72B, with the latter becoming the first open-source LLM to surpass an average accuracy of 80% on the HuggingFace Open LLM Leaderboard.

C2weakest assumption

That the identified DPO failure mode is the primary limiter of performance and that the DPOP modification will generalize without introducing new unintended effects on model behavior or other tasks.

C3one line summary

DPOP is a new loss function that prevents DPO from lowering preferred response likelihoods and outperforms standard DPO on diverse datasets, MT-Bench, and enables Smaug-72B to exceed 80% on the Open LLM Leaderboard.

References

288 extracted · 288 resolved · 31 Pith anchors

[1] 01.AI. Yi-34b-200k, 2024. URL https://huggingface.co/01-ai/Yi-34B-200K 2024
[2] Ultrafeedback binarized clean, 2024 2024
[3] Learning from mistakes makes llm better reasoner 2023
[4] A general theoretical paradigm to understand learning from human preferences 2023
[6] Training a helpful and harmless assistant with reinforcement learning from human feedback, 2022 2022

Formal links

3 machine-checked theorem links

Cited by

16 papers in Pith

Receipt and verification
First computed2026-05-17T23:38:12.705233Z
Builderpith-number-builder-2026-05-17-v1
SignaturePith Ed25519 (pith-v1-2026-05) · public key
Schemapith-number/v1.0

Canonical hash

6d1c0b2665de3792d4c940131711a151ec0949739234f6c31e62289f1583e999

Aliases

arxiv: 2402.13228 · arxiv_version: 2402.13228v2 · doi: 10.48550/arxiv.2402.13228
Agent API
Verify this Pith Number yourself
curl -sH 'Accept: application/ld+json' https://pith.science/pith/NUOAWJTF3Y3ZFVGJIAJROENBKH \
  | jq -c '.canonical_record' \
  | python3 -c "import sys,json,hashlib; b=json.dumps(json.loads(sys.stdin.read()), sort_keys=True, separators=(',',':'), ensure_ascii=False).encode(); print(hashlib.sha256(b).hexdigest())"
# expect: 6d1c0b2665de3792d4c940131711a151ec0949739234f6c31e62289f1583e999
Canonical record JSON
{
  "metadata": {
    "abstract_canon_sha256": "daca5ab3b31f5b6cbc9dd341d2cc0a479fe7a160a5d5c8d1ee44635c801b0169",
    "cross_cats_sorted": [
      "cs.AI",
      "cs.LG"
    ],
    "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/",
    "primary_cat": "cs.CL",
    "submitted_at": "2024-02-20T18:42:34Z",
    "title_canon_sha256": "1ec8c5d2a38e1c0b3274b675b8fd70cfced7d2cf0a2b756023af35a6b491730b"
  },
  "schema_version": "1.0",
  "source": {
    "id": "2402.13228",
    "kind": "arxiv",
    "version": 2
  }
}