Today
-
code: LoRA strength sweep
Yesterday
-
code: coding wipe family + single-GPU phases
-
code: robust wipe parsing
-
code: eval_AB phase
-
gen-A/gen-B with 121 DAgger rows (cutoff)
-
generalisation data gen-A/gen-B (pre-DAgger)
-
code: reviewed transfer eval + pair checks
-
code: transfer eval harness
-
code: DAgger rollouts + wipe families
-
code: kaggle grpo installs trl+agenttune
-
code: kaggle grpo5 pair
-
SFT v6 data (element-wise switch comparison, late-room replay) + code
-
SFT v5 data (explicit exits, switch diffs) + code
-
GRPO v4 states (rooms 1-3 + unlock) + code (T4 fp16 GRPO pair)
-
SFT v4 data (every arithmetic step explicit, door derivation at every bank) + code
-
SFT v3 data (evidence gate + room 1-3 derivations, balanced registry bit) + code
-
SFT v3 data (room 1-3 inference derivations) + code (modal sft/publish ops)
-
code: own-notebook chain eval
-
code: held-out step-reward metric
-
code: modal ops queue + eval token budget
-
code: GRPO checkpoints every 50 steps
-
GRPO v2 states + code (completion_len/temperature fix, modal runner)
-
code: lenient arm, run_specs
-
SFT v2 data (derivations) + code 9a9943de7f6e + job
-
bundle: notebook interventions
-
bundle: notebook interventions
-
bundle: fp16 on T4, JSON stop, in-job check
-
bundle: fp16 on T4, JSON stop, in-job check
-
bundle: hardened eval + preflight
-
bundle: grpo T4 memory settings
-
bundle: current code
2 days ago
-
Upload bundle/MANIFEST.sha256.txt with huggingface_hub
-
training bundle (text chunks)