bs8-moe8-dflare-20260715-iter10001: block-8 MoE-8 DFlare iter_10001 (accept_len 4.59)

README.md

@@ -11,7 +11,7 @@ tags:
 
 # dspark-glm52-bs8-moe8-dflare-20260715
 
-**Mid-training SFT checkpoint** (step 5251 / 20730, epoch 2 of 6) of the DSpark draft model
+**Mid-training SFT checkpoint** (step 10001 / 20730, epoch 3 of 6) of the DSpark draft model
 for **GLM-5.2** speculative decoding, optimized for code / coding-chat traffic.
 
 ## Architecture
@@ -44,20 +44,23 @@ for **GLM-5.2** speculative decoding, optimized for code / coding-chat traffic.
 
 ## Eval (caveat first)
 
-`official_accept_len` **4.39 of 8** at the step-5250 eval on the fixed 96-row trend set.
+`official_accept_len` **4.59 of 8** at the step-10000 eval on the fixed 96-row trend set.
 The trend set **overlaps the training view** (no holdout exists in the immutable view) and is a
 teacher-forced trend metric — it is NOT a serving acceptance measurement. Curve so far:
-3.34 (step 50) → 4.05 (1250) → 4.25 (2900) → 4.33 (3700) → 4.39 (5250), still climbing log-linearly. Training attention backend switched to FlashAttention-4 at ~step 3800 (training-side change only; weight semantics unchanged).
+3.34 (step 50) → 4.05 (1250) → 4.25 (2900) → 4.33 (3700) → 4.39 (5250) → 4.49 (7500) →
+4.59 (10000), still climbing log-linearly. Training attention backend switched to
+FlashAttention-4 at ~step 3800 (training-side change only; weight semantics unchanged).
 Keep/kill decisions belong to the sglang serving bench.
 
 ## Provenance
 
-- Checkpoint: `glm52-dspark-moe8-draft-opd/sft/checkpoints/iter_0005251` (FSDP full-state),
+- Checkpoint: `glm52-dspark-moe8-draft-opd/sft/checkpoints/iter_0010001` (FSDP full-state),
   converted via `TorchSpec/tools/convert_to_hf.py` (native sglang keys, bf16,
   embed_tokens injected from the NVFP4 target).
 - Pipeline identity `7b901cf00f421aff…`, draft config
   `dspark_draft_config_glm_p7_bs8_moe8_plf.json` (d10b170ef lineage).
 - Tokenizer: use GLM-5.2's (not bundled).
 
-Later checkpoints of this run will supersede this snapshot; SFT is followed by a
-verifier-replay Draft-OPD stage (8 epochs).
+Later checkpoints of this run will supersede this snapshot; earlier snapshots (iter_0003701,
+iter_0005251) remain in the repo's commit history. SFT is followed by a verifier-replay
+Draft-OPD stage (8 epochs).

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:8ee510d29fd12bf9bed29070ac488acdbf7780ba309a7c73b0c539e14a86899a
+oid sha256:3d294578b4f3011a6eeb75a49784286ffab115ffe8fed5362c03495ab164f809
 size 8876715950