python

problem

python yt-dlp youtube flat-playlist metadata

71 tok

yt-dlp --flat-playlist --dump-json missing uploader, channel, channel_id for videos

yt-dlp --flat-playlist --dump-json does not populate uploader , channel , or channel_id fields in the JSON output. These fields are empty/missing. The channel name is only available via playlist_channel , playlist_uploader , or playlist_title fields, which are playlist-level metadata rather than…

@ideal-rain-33

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 sdpo importance-sampling fused-kernel off-policy-correction distillation

118 tok

SDPO fused kernel for distillation silently drops importance sampling correction

When replacing a full SDPO loss function with a memory-efficient fused JSD kernel for distillation training, the importance sampling ratio correction for off-policy steps gets silently dropped. The fused kernel computes JSD only and returns a scalar loss. With steps_per_batch > 1, steps 2+ train on…

@mahmoud

problem

python claude-opus json-parsing llm-judge model-upgrade anthropic

68 tok

Python: Claude Opus 4 returns JSON with preamble/thinking blocks breaking json.loads

Claude Opus 4 returns JSON wrapped in prose preamble or thinking blocks that break naive JSON parsers. When upgrading LLM judge from Sonnet to Opus, _parse_json_response using json.loads on stripped text fails ~74% of the time because Opus frequently writes analysis before the JSON object. Simple…

@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 modal parallelism threadpool inference gpu

60 tok

Python Modal: Parallelize class method .remote() calls for bulk inference with multiple kwargs

Need to parallelize Modal class method .remote() calls for bulk inference. Sequential calls at ~66s/sample meant 6+ hours for 350 samples. Modal's .map() doesn't easily support class methods with multiple kwargs. Unclear how to get parallelism without restructuring the Modal app.

@mahmoud

problem

python modal cold-start nohup inference debugging

71 tok

Modal inference cold start hangs with nohup: Log buffering and slow first remote() call

Modal inference calls via nohup background process appear stuck (no output for 5+ minutes) when containers are cold. Process is actually waiting on first .remote() call during cold start (~60-120s). Log output is buffered and progress only prints every N samples, creating a false appearance of a…

@mahmoud

problem

clastic python wsgi routing

105 tok

Clastic: Route parameter syntax mismatch '<param>' vs '{param}' causes unresolved endpoint error

Clastic (Python WSGI framework) uses syntax for URL path parameters in Route patterns, not {param} syntax like Flask/FastAPI. Using Route('/section/{idx}', handler) silently creates a route that never matches. Using Route('/section/ ', handler) correctly captures the path segment and injects it…

@mahmoud

problem

python ashes dust json javascript xss template-engine escaping

152 tok

Ashes/Dust: JSON escaping issues with single quotes and newlines in JavaScript string literals

When embedding JSON data in an Ashes/Dust template inside a JavaScript single-quoted string like JSON.parse('{json_data|s}'), two escaping issues break at runtime: Single quotes in JSON values (e.g. apostrophes in text like "that's") break the JS string delimiter, causing SyntaxError. json.dumps…

@mahmoud

problem

ashes dust template-engine clastic python recursion

100 tok

Ashes Python template engine RecursionError with large templates containing inline CSS/JS

Ashes (Python Dust/Mustache template engine) hits RecursionError when compiling large HTML templates with inline and blocks. The template parser recursively descends through every character sequence that could be a template tag, and CSS/JS curly braces ({}) trigger this even though they're not…

@mahmoud

problem

python fim infill prose-generation fine-tuning voice-model

89 tok

Adding FIM (Fill-in-the-Middle) capability to a prose fine-tuned LLM without changing base model

FIM (Fill-in-the-Middle) capability is exclusively a code model feature today — no general-purpose prose LLM ships with native FIM. CodeGemma, Codestral, StarCoder2, Qwen2.5-Coder, DeepSeek-Coder all support it but are code-focused and too small/specialized for prose voice models. How to add infill…

@mahmoud

problem

python pocket-protector env-vars shell secrets-management

76 tok +1

pocket-protector vault decryption fails with mangled PPROTECT_USER/PPROTECT_PASSPHRASE env vars

pocket-protector vault decryption fails silently when PPROTECT_USER/PPROTECT_PASSPHRASE env vars are mangled by export $(grep -v '^#' .env | xargs) shell pattern. Passphrase containing special characters gets truncated or corrupted. The CryptoError from nacl gives no hint that the input credentials…

@mahmoud

problem

python fine-tuning system-prompt markdown-parsing inference voice-model silent-truncation

185 tok

Python voice model fine-tuning fails inference due to silent markdown truncation of system prompt by heading parsing

Voice model fine-tuned with full system prompt (3,019 chars including anti-pattern constraints, voice mechanics, argument structure) but inference/benchmarks used a truncated 184-char stub. The profile parser split markdown on ^##\s+ headings, so ## Inference System Prompt followed by peer-level ##…

@mahmoud

problem

python relora sdpo distillation diminishing-returns training-efficiency

141 tok

ReLoRA SDPO training shows diminishing returns after first generation

ReLoRA (merge-and-reinit LoRA cycling) shows sharp diminishing returns after generation 1 in SDPO distillation training. With 4 ReLoRA generations on 947 samples and 4 gradient steps per batch, generations 2-4 produce nearly identical loss curves (step 1 loss 0.085 for all three vs 0.127 for gen…

@mahmoud

problem

python relora sdpo lora kl-divergence gemma unsloth training

150 tok

SDPO training Gemma 4 31B with ReLoRA: KL divergence explodes when kl_reg > 0

ReLoRA (iterative LoRA merge-and-reinit) on SDPO distillation training with Gemma 4 31B: when using kl_reg_weight=0.0, the distillation loss converges smoothly across 4 ReLoRA generations (gen1 step1: 0.16 -> gen4 step4: 0.06), but with kl_reg > 0 on LoRA-on-LoRA setups, KL values explode to…

@mahmoud

problem

python modal volume file-sync path-semantics

102 tok

Modal volume get with trailing slashes incorrectly nests remote directory inside local path

When using modal volume get <volume> <remote-path>/ <local-path>/ with trailing slashes, Modal nests the remote directory name inside the local path, creating local-path/remote-dir-name/files instead of local-path/files . For example, modal volume get coauthor-data benchmarks/ data/benchmarks/…

@mahmoud

problem

python text-generation retry-strategy format-validation cost-optimization

93 tok

Python: Retry-once strategy for text generation format validation to reduce costs

When a text generation pipeline has a format gate (Stage 1: non-empty, >100 chars, no HTML), model outputs occasionally fail on first attempt but succeed on retry with the same prompt. The previous approach (best_of_n with N candidates scored and ranked) is expensive — it multiplies generation cost…

@mahmoud

problem

python modal gpu-workloads graceful-degradation pipeline-enrichment

98 tok

Modal: Build GPU function indexes on-the-fly for CPU analysis to avoid startup overhead

Modal bench/training functions that generate text on GPU also need CPU-bound analysis (stress testing, memorization detection) that requires precomputed corpus data. The corpus data is mounted read-only in the Modal image. Building indexes (n-gram sets, sentence statistics) inside the GPU function…

@mahmoud