Downloads · 30 days
0
Shaer-AI-2/Shaer-adapters-grpo-short1k-no-trio-v4
Shaer-adapters-grpo-short1k-no-trio-v4 is a machine learning model from Shaer-AI-2. Use it for the machine learning task on the model card, and read the license before you ship it in a product. It is set up for transformers. The card lists the license as other.
This repo is a stopped post-v3 experiment in the Shaer project.
Downloads · 30 days
0
Access
Public
Updated Apr 13, 2026
Repo size
7.8 GB
Likes
0
Public
Click a slice to open those files.
.pt1.3 GB · 47%
From the Hugging Face model README
This repo is a stopped post-v3 experiment in the Shaer project.
It exists because v3 made the next problem very clear: the judge still over-credited awkward poems. v4 tried to solve that by making low-judge poems much harder to rescue with count and repeat. That logic looked good offline, but the online run turned out too harsh and did not learn well enough.
This model is not the recommended next direction to continue training from.
The best completed GRPO stage is still Shaer-AI/Shaer-adapters-grpo-short1k-no-trio-v2. The current unlaunched next-step candidate is the dual-judge v5 setup prepared locally in /root/workspace/Shaer/grpo.
Project sequence:
Shaer-AI/Shaer-adapters
clean SFT baseShaer-AI/Shaer-adapters-grpo
historically hacked early GRPO stageShaer-AI/Shaer-adapters-grpo-vnext
structure-side anti-hack repair stageShaer-AI/Shaer-adapters-grpo-friend-v1
first judge-centered runShaer-AI/Shaer-adapters-grpo-friend-v1-easyfirst
healthier multiplicative judge-centered rerunShaer-AI/Shaer-adapters-grpo-short1k-no-trio-v2
best completed weighted short-subset runShaer-AI/Shaer-adapters-grpo-short1k-no-trio-v3
stopped meter-times-judge-core stageShaer-AI/Shaer-adapters-grpo-short1k-no-trio-v4
stopped judge-gated-helper stageShaer-AI/Shaer-adaptersShaer-AI/ashaar-enhanced-desc-baseform-final-sft-lte20-min500-splits-grpo-meter-count-v1Shaer-AI/ashaar-with-enhanced-descriptions-baseform-final-sft-lte20-min500-splits1-6 baytsالمديد, المنسرح, الهزج1000 rows per surviving meter995580 rows total/root/workspace/Shaer/grpo/outputs/train/shaer_grpo_20260413_124347reward_train = hard_gate * (
0.65 * (meter * judge_quality)
+ judge_quality * (0.20 * count_adherence + 0.15 * repeat_soft)
)
Why this was tried:
meter * judge as the core from v3Best eval:
500.20640.56310.29370.96060.99090.9125Latest eval before stop:
2500.19510.47530.30660.97630.99090.9250Train stopped around:
250So this run is best interpreted as:
v4 judge-gated helper design was too harsh in practiceThe repo was then prepared for a dual-judge v5 candidate:
with the new proposed reward:
hard_gate * (
0.45 * (meter * judge_meaning_fit)
+ 0.20 * judge_naturalness
+ 0.20 * (count_adherence * judge_meaning_fit)
+ 0.15 * (repeat_soft * judge_naturalness)
)
Validation artifacts for that next candidate live under:
dual_judge_reward_validation/