-
Option-B throughput: gradient checkpointing off, dev eval subsampled to 300, eval/save every 200 steps
-
Add push_with_retry (transient Hub 500s), import time
-
Fix push_with_retry call (missing out_dir arg)
-
Retry push_to_hub (transient Hub 500s); report local-only fallback instead of dying
-
Add --epochs with dynamic warmup (option B: 1-epoch runs)
-
Add --epochs with dynamic warmup (option B: 1-epoch runs)
-
Fix tokenizer source (args.model, not nonexistent args.model_id)
-
New: constrained-decoding inference over test queries -> preds.json for eval_retrieval.py
-
Fix SFTTrainer model= kwarg, wire dev eval dataset, fix warmup for 2930-example train set, trackio env-gated
-
Fix SFTTrainer model= kwarg, wire dev eval dataset, fix warmup for 2930-example train set, trackio env-gated
-
Fix out_dirs bug: write questions.jsonl to GEN_ROOT per-book dirs, not source dirs
-
Env-driven teacher (unquantized Mixtral, TP via env, GPTQ fallback), dual-format smoke gate
-
Add pre-generation smoke gate: abort if 0/5 smoke prompts parse
-
Harden teacher-response parsing: max_tokens 160, raw dump, robust extraction
-
Upload gen_questions.py with huggingface_hub
-
Upload gen_questions.py with huggingface_hub
-
Upload eval_retrieval.py with huggingface_hub
-
Upload train_dsi.py with huggingface_hub
-
Upload train_stair.py with huggingface_hub
-
Upload stair_common.py with huggingface_hub
-
Upload gen_questions.py with huggingface_hub
-
initial commit
