Downloads · 30 days
0
Zhongzhu/tunekv-mcpuniverse-s1
tunekv-mcpuniverse-s1 is a machine learning model from Zhongzhu. Use it for the machine learning task on the model card, and read the license before you ship it in a product.
multi-swe §8 MCP-Universe, s1-maintenance slice (v3-keyless-judge subset, SHA 6a0b6c17, 50 tasks × 4 rounds, temp 1.0): 30B GRPO outcome-RL arm, DSV4-Flash base-uniform50 rerun (old caliber), 27B-base v3 completion.
Downloads · 30 days
0
Access
Public
Updated Sep 24, 2026
Repo size
1.6 GB
Likes
0
Public
Click a slice to open those files.
.tgz1.7 GB · 96%
From the Hugging Face model README
multi-swe §8 MCP-Universe, s1-maintenance slice (v3-keyless-judge subset,
SHA 6a0b6c17, 50 tasks × 4 rounds, temp 1.0): 30B GRPO outcome-RL arm,
DSV4-Flash base-uniform50 rerun (old caliber), 27B-base v3 completion.
All v3-subset numbers are NOT comparable to the v2-subset rows in PLAN §8 (V1–V16 etc.); within-slice paired comparisons only.
Sibling of the s1-owner V6-CE arm (seed-10 rows): single variable = loss + data (10 gold rows → 232 kept K=8 train rollouts, per-task z-scored). Prefix: owner's 1103-token preamble-headed span, n_tune 1024, capture init. Rollout: 37 key-free train tasks (fin 19 / br 9 / 3d 9) × K=8 = 296 rounds, temp 1.0 WITH preamble, TP2 B300 base serve; 15 successes (d3 9 / br 6 / fin 0). Kept 232 rows (4 crash-drop, 26 no-transcript, 34 too_long >32768); 3 mixed groups / 23 nonzero-adv rows (d3_0011 5/8, d3_0019 4/8, br_sports_0004 6/7); 121 sequential bins → 30 steps (dropped bin = 3 all-zero fin rows, zero signal lost). Train: 1×B300, xorl, 30/30 steps, admission all-green (96 trainable all changed, 435 frozen 0 changed), lr 1.5e-3. Eval (same TP2 stack, WITH preamble): tuned 11/200 = 5.5%±1.91 (rounds 4/3/2/2; fin 0/br 9/3d 2) — LP A/B vs base: mean|dLP| 2e-4 over 32 tokens (KV active, effect tiny). base_rt (same stack + preamble): 10/200 = 5.0%±2.00 (rounds 3/1/3/3; fin 0/br 8/3d 2) → tuned−base Δ+0.5pp, GRPO true null (not a preamble effect).
kv_grpo30.tgz — trained prefix_kv (global_step_30, TP2-split, 96 shards + SHA.json + kv_transfer.json)rows_grpo.jsonl / manifest_grpo.jsonl — 232 training rows + per-row reward/advantageTRAINING.json / train_grpo30.log — admission evidence (30/30 steps)arm_mcpu30_grpo.yaml — exact config (max_steps 30 derived from packing census)score_grpo_tuned.json — 200-round eval scoreseval_grpo_tuned.tgz — full trajectorieseval_grpo_base_rt.tgz / score_grpo_base_rt.json — same-stack base control (10/200 = 5.0%)lp_tuned_probe.json / lp_base_probe.json — LP A/B fingerprintsprefix_mcpu_coder30b.json / prefix_ids_list.json / train_pool_37.txt / subset_50.txtwandb: zhoumllab/mcp-universe run m8s1-30b-grpo-outcome-zscore
(3acfhglr; first attempt kqz9ujpn crashed on dict-vs-list prefix format).
Full-fresh 50×4 (the 09-21 6/50 r1 + owner r2–4 are void: max_tokens 10000,
no effort prompt). Weights deepseek-ai/DeepSeek-V4-Flash@60d8d707
(46 shards, 149G); serve = 09-21 replica (vLLM 0.28 TP2, fp8 KV, 131072,
deepseek_v4 parser, 2×B300); harness = official chat path + weights'
own REASONING_EFFORT_MAX text prepended (byte-verified against
encoding/encoding_dsv4.py in the weights — the "Absolute maximum" tier,
NOT the newer vendored "Beyond maximum" tier) + --max-tokens 32768.
Result: 32/200 = 16.0%±1.63 (rounds 8/9/7/8; fin 9/br 16/3d 7).
score_dsv4_rerun.json, eval_dsv4_rerun.tgz, serve_dsv4.sh, subset_50.txtOwner queue had 197 valid + 3 stale-claimed; 3 resume rounds on owner's
charlie serve closed it to 200/200 (3 truncated lines skipped, all covered
by dedup). 44/200 = 22.0%±0.0 (rounds 11/11/11/11; fin 11/br 25/3d 8).
Raw trajectories live in the owner dir
(work/confidential-maintain-mcpu8/data/eval_27b_base/); only the score
is staged here.
Owner's 27B V6 training (final_only × 0.3 reverse KL, 50/50 steps, CE→0.04) adopted as-is (wandb 4l572y63); eval run 09-23 14:27Z→19:26Z on 009 dual TP1 replicas (8161 reused 17h+, 8162 fresh on GPU0, same kv_tuned50, replica-agreement probe passed), preamble + react, same harness/subset as base-v3. 95/200 = 47.5%±5.97 (rounds 25/20/23/27; fin 46/br 36/3d 13) vs base-v3 22.0%: Δ+25.5pp, McNemar 40.0 (58 gained / 7 lost, p<1e-9). Disclosure: base ran on charlie serves, CE on 009 TP1 serves (same software/flags per owner docs); train manifest eval_leak=0.
kv_tuned50.tgz — assembled tuned prefix-KV (67 files: 16 FA + 48 GDN states + COMPOSITION.json)eval_27b_ce_v6.tgz — full trajectories (200/200, 0 missing)score_27b_ce_v6.json / paired_27b.json — arm score + paired flipspreamble_qwen38_27b.txt / prefix_mcpu_27b_serve.json / subset_50.txt — protocol