GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
wetheridge
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
wetheridge
Model Tree
Qwen/Qwen2.5-1.5B-Instruct
Input Modalities
Output Modalities
Supported Functionality
The adapter is intended to be loaded on Qwen/Qwen2.5-1.5B-Instruct at revision 989aa7980e4cf806f80c7fef2b1adb7bc71aa306 and invoked with the model's normal chat template and an empty system message.
The desired response contains exactly:
The exact response schema and full research process are available in the ChatG&T source repository.
Qwen/Qwen2.5-1.5B-Instructq_proj, k_proj, v_proj, o_proj2e-4202607151.966936, down from a 3.036822 baselineThe checkpoint was selected by the lowest validation loss within its own fixed training trajectory. Held-out prompts were not used for training, checkpoint selection, or tuning.
The frozen evaluation used 60 unseen prompts and identical sampled-generation settings for all systems.
| System | Schema-valid JSON | Full response pass |
|---|---|---|
| Base model + five-shot prompt | 43/60 (71.7%) | 16/60 (26.7%) |
| This adapter + empty system prompt | 52/60 (86.7%) | 26/60 (43.3%) |
The adapter improved schema validity by 15 percentage points while reducing mean rendered input length from 2,575.9 tokens to 31.9 tokens. In the matched local measurement it was also 1.78 seconds slower on average, so prompt-token reduction did not translate directly into lower model-generation latency.
Qualitative judgments were less certain than the structural result. A sealed 15-pair human calibration agreed with the primary LLM pairwise judge on 7/15 comparisons. The strongest supported conclusion is therefore that the adapter improved structured-output reliability; broad claims about usefulness or preference should remain qualified.
python
from peft import PeftModelfrom transformers import AutoModelForCausalLM, AutoTokenizerbase_id = "Qwen/Qwen2.5-1.5B-Instruct"adapter_id = "wetheridge/chatgnt-qwen2.5-1.5b-lora"tokenizer = AutoTokenizer.from_pretrained(base_id)base = AutoModelForCausalLM.from_pretrained(base_id, dtype="auto")model = PeftModel.from_pretrained(base, adapter_id)
Use the base model's chat template with an empty system message. Generation settings used in the experiment were temperature 0.7, top-p 0.8, top-k 20, repetition penalty 1.1, and at most 512 new tokens.
0eef1d5da017a18f571a5e37cc152d1351bfbda9f0d343c3f376a1f68a2d4249034e0c79b1201350784e0409d5f28ffcf92d8dfa9697b291b2dd4bf45129100cdd924a11b4c220f385b51ffa522daea7c9f3d850e31b162bb5661df483c6d3eeThe published package changes only deployment metadata in adapter_config.json: the training machine's local base-model path is replaced with the canonical Hub model ID and pinned revision. The safetensors weights are byte-identical to the evaluated training artefact. Both identities are recorded in publication-provenance.json.