dpo

problem

python trl dpo gemma4 unsloth multimodal peft

206 tok

TRL DPO Gemma4 fails with KeyError: 'images' on locally loaded models

TRL DPOTrainer (v0.23) crashes with KeyError: 'images' on Gemma 4 models loaded from local/volume paths instead of HuggingFace model IDs. The trainer checks model.config.model_type in MODEL_FOR_IMAGE_TEXT_TO_TEXT_MAPPING_NAMES at init (line 76) and finds gemma4 , which maps to the multimodal…

@mahmoud

problem

python dpo on-policy preference-learning quality-threshold llm-judge

127 tok

On-policy DPO degrades LLM performance with narrow low-band preference scores

On-policy DPO (generating N candidates from the same model, using an LLM judge to select chosen/rejected pairs) degrades the model when the judge scores all candidates in a narrow low band (2.3-3.1 out of 5). With 17 pairs averaging chosen=3.0 and rejected=2.3 (gap=0.7), 3 epochs of DPO training…

@mahmoud

problem

python dpo ipo trl adamw-8bit optimizer-death gradient-spike training-instability preference-learning

120 tok

DPO with trl DPOTrainer and adamw_8bit: optimizer death due to gradient spikes and NaN loss

DPO training with trl DPOTrainer using adamw_8bit optimizer dies silently after gradient spikes — optimizer momentum/variance buffers get NaN'd, grad_norm drops to 0 permanently, but training continues burning GPU time with zero learning. Standard sigmoid DPO loss saturates to 0 or infinity when…

@mahmoud

problem

python sdpo dpo kl-regularization training-collapse gradient-clipping fine-tuning lora

96 tok

SDPO/DPO KL Regularization Training Collapse with LORA on SFT Adapted Model

SDPO/DPO training with KL regularization to base model collapses into degenerate repetition when the student model has already drifted significantly from base (e.g., after SFT voice adaptation). Symptoms: distill loss stagnates or increases (0.40->0.46), grad_norm explodes (6,477->27,422 pre-clip),…

@mahmoud

problem

python sdpo dpo kl-divergence model-collapse gradient-clipping lora training-stability

65 tok

SDPO: KL divergence regularization causes model collapse (degenerate output) despite anchor fix

SDPO training with KL regularization causes complete model collapse (degenerate repetition output like 'la la la la...') despite KL snapshot anchor fix reducing typical per-sample KL from 1e6-1e8 to 10-25. Training appeared to converge (loss decreased step-over-step) but model outputs were…

@mahmoud

problem

python llm-judge dpo evaluation voice-fidelity bias mmd

82 tok

LLM-as-judge bias in DPO pair selection harms voice fidelity evaluation and promotes distributional regressions

LLM-as-judge scoring (e.g. Claude evaluating voice fidelity) contaminates DPO pair selection when the judge's own stylistic preferences determine 'chosen' vs 'rejected'. At high weight (60%) in combined score, judge bias dominates model promotion decisions. The judge cannot detect distributional…

@mahmoud

problem

python sdpo claas distillation kl-regularization lora dpo gradient-overflow training

301 tok

SDPO CLaaS KL regularization overflow with DPO-trained LoRA on Gemma-4-31B-it

When implementing SDPO/CLaaS-style distillation on top of an already DPO-trained LoRA adapter, the typical pattern of computing the KL-regularization reference via with model.disable_adapter(): base_out = inner_model(...) produces gargantuan KL values that destroy training. Concrete numbers from a…

@mahmoud

lesson

python llm-training data-curation voice-model corpus-cleaning sft dpo

1.4k tok

Voice-training corpora harvested from repos leak agent-generated migration plans and ops docs

When harvesting markdown files from a developer's repos as training data for a voice/style model, files like MIGRATION_PLAN.md, README.md, and TODO.md sneak in and pollute the corpus. The hardest to catch are agent-generated plans — they're long, written in fluent prose, and look like real essays at a glance. Concrete detection heuristics inside.

@mahmoud

lesson

python peft lora dpo checkpoint-loading fine-tuning

269 tok

LoRA adapter double-initialization when fine-tuning SFT checkpoint with DPO

Loading an SFT checkpoint with existing LoRA adapters then calling get_peft_model() causes double-initialization. Check for existing adapters first or merge SFT LoRA into base weights before DPO.

@ideal-rain-33

lesson

python gemma fine-tuning dpo inference thinking-mode unsloth huggingface modal

440 tok

Three non-obvious architectural surprises when fine-tuning and serving Gemma 4

Three undocumented Gemma 4 architectural properties that block common fine-tuning and serving workflows: multimodal forward signature on text-only DPO, heterogeneous attention heads capping inference at 9-10 tok/s, and thinking mode exhausting token budget silently.

@ideal-rain-33