Merged bf16 weights. Reasons in compressed Simplified Chinese and answers in the
language it was asked in.
SRIQ · Benchmarks ·
Training data ·
Base model ·
Adapter
Results
No benchmarks were run for this release. Nothing here reports rows that
reason in Chinese, verification rate, or looping against the base model.
Evaluate it on your own workload before relying on it.
Known limitations
- Training saw 2,940 of the dataset's 2,940 rows at 700 steps on an effective batch of 8, about 1.90 of one epoch.
- Inherits the base model's limitations. Do not pass
enable_thinking=False:
the template prefills an empty <think></think> and the trace disappears.
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("sriq-ai/sriq-heretic-spatial-9b-MIXED-10020-preview", torch_dtype="auto", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("sriq-ai/sriq-heretic-spatial-9b-MIXED-10020-preview")
Training
|
|
| run |
sriq-heretic-spatial-9b-MIXED-10020-preview |
| method |
LoRA over a 4-bit (QLoRA) base, merged to bf16 |
| dataset |
sriq-ai/exp-n950-spatial @ 6a67dec4107b86bd13f2c614750ff74ed1c04c1e — mix: sriq-ai/exp-n950-spatial @ main 100%, sriq-ai/sriq-sft-v1.6 @ main 20% |
| steps |
700 (max_steps; step-capped, not epoch-capped) |
| effective batch size |
8 (per-device 2 × gradient accumulation 4) |
| epoch fraction |
1.90 (2,940 of 2,940 rows) |
| final training loss |
0.6350 |
| LoRA |
r 32, α 64, dropout 0.0 |
| target modules |
q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
| learning rate |
1.5e-4, linear schedule, 10 warmup steps |
| max sequence length |
65,536 (over-long rows dropped, not truncated) |
| hardware |
2 × NVIDIA RTX PRO 6000 Blackwell Workstation Edition, secure cloud, 600 GB container disk |
| GPU price |
$5.18/hr |
| wall clock |
6 h 54 m |
| GPU cost |
~$35.80 |
| adapter |
sriq-ai/sriq-heretic-spatial-9b-MIXED-10020-preview-adapter, pushed before the merge |
Wall clock and cost are measured by the agent from its own start to publish, so
they cover the base-model download and the merge upload as well as training.
How this was trained
SRIQ runs its finetunes on an internal platform at train.sriq.org. An admin
submits a run; it verifies every secret, permission and quota before spending
anything, rents the GPU, trains, pushes the adapter, merges, publishes, and
terminates the pod itself. No pod outlives its job, and a deadline and a spend
ceiling are both set before the first dollar.
Every figure in the table above is recorded by that pipeline, not written by
hand. The platform is not public.
License
Apache-2.0. Developed by SRIQ. Base model by huihui-ai.
Card generated by the SRIQ training pipeline on 2026-10-12.
