69 lines
2.5 KiB
Markdown
69 lines
2.5 KiB
Markdown
---
|
|
license: apache-2.0
|
|
library_name: transformers
|
|
pipeline_tag: text-generation
|
|
base_model: Qwen/Qwen3-1.7B-Base
|
|
---
|
|
|
|
# Vikra-HCT-YeAM-3_3.2_QweLLa-1.7B
|
|
|
|
HCT architecture release. YeAM (Yet Another Merge) implementation invariant.
|
|
|
|
## What it is
|
|
|
|
A compact 1.7B-class checkpoint produced via HCT-compatible merging.
|
|
The checkpoint is published in standard Hugging Face format (safetensors + index).
|
|
|
|
## YeAM summary
|
|
|
|
YeAM performs a controlled merge in a real 4D geometric formulation with ray-intersection alignment in parameter space.
|
|
It also supports targeted knowledge injection (distillation-style) into a chosen model while remaining HF-compatible.
|
|
|
|
## Notes for this checkpoint
|
|
|
|
Compared to other YeAM/HCT merges, this checkpoint additionally applies a targeted merge on Attention projection weights.
|
|
|
|
Observed behavior tends to include characteristic Llama-like traits:
|
|
- More Llama-style conversation patterns.
|
|
- More consistent formatting.
|
|
- Stronger RLHF-like refusal/priority behaviors.
|
|
- Reasoning / chain-of-thought style output in the model's full native format is expected to work.
|
|
|
|
At the same time, most Qwen3 behavior should theoretically remain, but due to knowledge/logic injection from the Llama side, some Qwen-specific properties may be partially degraded or inconsistent.
|
|
|
|
Repetition / looping:
|
|
- There is no universally perfect sampling configuration.
|
|
- At higher temperature, without a repetition-style penalty, the model may enter repetition loops.
|
|
- Pay special attention to repetition-related controls (e.g. repetition penalty / presence penalty) if you observe cycling.
|
|
|
|
Do not ask the model who created it.
|
|
In this specific merge, it may oscillate between incompatible parents (Alibaba vs Meta”), fail to settle, and get stuck in a sad loop.
|
|
|
|
## Usage (Transformers)
|
|
|
|
```python
|
|
from transformers import AutoTokenizer, AutoModelForCausalLM
|
|
import torch
|
|
|
|
m = "/path/to/Vikra-HCT-YeAM-3_3.2_QweLLa-1.7B"
|
|
|
|
tok = AutoTokenizer.from_pretrained(m, use_fast=True)
|
|
model = AutoModelForCausalLM.from_pretrained(
|
|
m,
|
|
torch_dtype=torch.bfloat16,
|
|
device_map="cuda",
|
|
).eval()
|
|
|
|
inputs = tok("Hello!", return_tensors="pt").to(model.device)
|
|
out = model.generate(**inputs, max_new_tokens=128)
|
|
print(tok.decode(out[0], skip_special_tokens=True))
|
|
```
|
|
|
|
## GGUF
|
|
|
|
Convert and quantize with llama.cpp (example):
|
|
|
|
```bash
|
|
python3 /path/to/llama.cpp/convert_hf_to_gguf.py /path/to/model --outtype f16 --outfile model.f16.gguf
|
|
/path/to/llama.cpp/build/bin/llama-quantize model.f16.gguf model.Q8_0.gguf Q8_0
|
|
``` |