Today
-
Match original Sharpen Qwen3-8B-Base future-mass LR and global batch to OPPD
Yesterday
-
Future-mass LR1e-6 speed variant bc77ed7 immutable snapshot
2 days ago
-
Snapshot full future-mass OPPD beta1 global-particle training 865707e
-
Sharpen-style MH monitoring metrics and independent CPU rollout grading
-
Use installed SwanLab SDK URL property and resume monitoring run
-
SwanLab monitoring adapter for running OPPD MH experiment
-
Snapshot OPPD MH-guided DDP with OPPD response-length normalization f9a6eef
3 days ago
-
Pin unmodified official OPPD 784a7fd for Qwen3-8B reproduction
-
Keep snapshot identity in manifest root and synchronization metadata
-
Release full-parameter actor gradients before future-mass OPPD rollout
4 days ago
-
Use future mass in an official OPPD block teacher and train Qwen3-8B-Base
-
Use future mass in an official OPPD block teacher and train Qwen3-8B-Base
-
Run alpha-annealed future-guided OPPD on Qwen3-4B-Base with consistent plain prompts
-
Record final training particle accuracy in the existing SwanLab run
-
Align guided OPPD block FKL and add official boxed-answer stopping
5 days ago
-
Use FP32 master weights with BF16 compute for full-parameter OPPD
-
Fix full-parameter frozen base initialization with typed rollout configuration
-
Enable full-parameter future-guided OPPD with independent frozen base engines
-
Fix future-guided vLLM sampler startup and restart on GPUs0-3
-
Fix future mass GPU probe server LoRA initialization
-
Configure future-guided OPPD for two GPUs, keeping batch32
-
Add isolated future-guided OPPD sampling-only experiment
-
Require matching particle trees and teacher evidence in execution benchmark
-
Complete execution probe EOS handling and confirmed8192 launcher check
-
Optimize equivalent future-mass OPPD execution and add timing probe
6 days ago
-
Set future-mass OPPD training generation cap to8192
-
Install future-mass OPPD full method for Qwen3-4B
7 days ago
-
Fix FSDP2 teacher initialization and add four-GPU anchor optimizer check
-
MH token anchor: exact token entropy and candidate weight concentration metrics
-
MH anchor: declare typed policy loss fields required by Hydra
-
MH anchor GPU check: compare identical BF16 forward shapes
-
MH power: tokenwise low-temperature forward-KL anchor; warmup10
8 days ago
-
Set MH evaluation frequency10 and checkpoint frequency25
-
Use token-budget dynamic batches for MH actor
-
Remove repeated future replay with native probability evidence
-
Use AIME24 and AIME25 only for MH power validation
-
Fix MH rollout custom configuration dictionary access
-
Add MH power distillation with B64 H63
10 days ago
-
Stage step100 validation config: 32768 response tokens
12 days ago
-
Enable centered-power vLLM prompt scoring and future hook at startup
-
Install centered power distillation for fresh Qwen3-4B training
13 days ago
-
Enable requested pre-training evaluation for optimized Sharpen restart
-
Honor disabled Sharpen validation frequency after successful one-step GPU speed test
-
Validated Sharpen future speed optimization; skip already completed baseline evaluation
-
Set explicit attention backend for deterministic Qwen3 weight verification
-
Record per-token future probability diagnostics before validating
-
Optimize exact power futures and block scheduling
-
Select the CUDA cross-entropy kernel only for CUDA tensors
-
Disable launch tracing of authentication values
-
Install ESS power Sharpen with corrected futures and single-pass actor
