GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
liuyihong
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
liuyihong
Model Tree
Qwen/Qwen3-VL-32B-Instruct
Input Modalities
Output Modalities
Supported Functionality
The archived SFT comparison is descriptive. The intended matched SFT-side run was incomplete, so this artifact must not be interpreted as a checkpoint-only causal estimate of the effect of SFT.
The corpus contains 10,000 example records derived from 5,000 paired tasks: 5,000 proposal examples and 5,000 Critic-conditioned repair examples. A 2% example-level evaluation split produced 9,800 training records and 200 evaluation records. A task-grouped split manifest was not retained, so the partitions are not claimed to be task-disjoint.
q_proj, k_proj, v_proj, o_proj, gate_proj,
up_proj, and down_proj;The training artifact did not retain an explicit Hugging Face revision for the
base checkpoint (revision=null). The base model identifier is therefore
reported without guessing a training-time revision. The later inference audit
used a local Qwen3-VL-32B-Instruct snapshot whose cache metadata recorded
revision 0cfaf48183f594c314753d30a4c4974bc75f3ccb; this does not by itself
prove that the same snapshot was used at the start of SFT.
adamw_torch), betas (0.9, 0.999), epsilon 1e-8;5e-5;The archived framework versions were PEFT 0.17.1, Transformers 4.57.6, PyTorch 2.5.1+cu124, Datasets 4.0.0, and Tokenizers 0.22.2.
The loss curves and JSON metric exports are included. training_config.yaml
contains the human-readable training configuration. Optimizer state,
checkpoints, RNG state, and Python-pickle training arguments are intentionally
excluded from this public adapter package.
The code and experiment configuration are maintained at https://github.com/RUCAIBox/VEGA-SR. The exact paper release tag and archival DOI should be cited together with the versioned adapter tag.
The release manifest is SHA256SUMS. In particular, the original adapter
weight artifact had SHA-256:
text
0a8b2c55edd3c9900a94aa55a0170979c748fd11b6e5fa88531c4fe88f693dbb adapter_model.safetensors
The base model and this LoRA adapter are distributed under the Apache License 2.0. A copy of the license is included in this model repository. Third-party software and data retain their respective upstream terms.
best_model_checkpointbest_metric