GLM4.7-Distill-LFM2.5-1.2B

Name: GLM4.7-Distill-LFM2.5-1.2B
Availability: InStock

Model Overview

GLM4.7-Distill-LFM2.5-1.2B is a 1.2B-parameter instruction-following language model obtained via offline distillation from GLM-4.7 into the Liquid AI LFM2 architecture.

The model is designed to be:

concise and non-verbose
strong at instruction following
efficient for local and edge deployments
suitable for assistant, agentic, and system-integration use cases

This model does not include chain-of-thought reasoning and is optimized for final-answer quality rather than verbose explanations.

Key Characteristics

Base architecture: Liquid AI LFM2
Model size: 1.2B parameters
Training method: Offline supervised distillation (SFT with LoRA)
Teacher model: GLM-4.7 (used only for data generation)
Inference dependency on teacher: None
Reasoning traces: Not included
Target behavior: Clear, grounded, instruction-aligned responses

Training Details

Distillation Approach

This model was trained using offline distillation, where instruction-response pairs generated by GLM-4.7 were combined with high-quality public instruction datasets.

The teacher model was not used during training or inference, and no teacher weights or logits are included.

Training focused on:

instruction adherence
response clarity
reduced verbosity
stable decision boundaries

Datasets Used (Approx. 13K Samples)

The following datasets were sampled and combined:

Open-Orca / FLAN
Databricks Dolly 15K
OpenAssistant OASST1
BAAI Infinity-Instruct
CodeAlpaca
TIGER-Lab MathInstruct

These were augmented with GLM-4.7–generated instruction responses, with explicit avoidance of chain-of-thought reasoning.

Intended Use

This model is well suited for:

general-purpose assistants
planning and task decomposition
summarization and explanation
lightweight coding assistance
agentic workflows
system integration and automation
on-device or edge inference scenarios

Limitations

Like other compact distilled models, this model may:

hallucinate when given insufficient or false premises
struggle with adversarial logical inference (NLI-style tasks)
lack temporal awareness of recent events
provide confident answers where explicit uncertainty is required

For critical reasoning, verification layers or external tools are recommended.

Ethical & Responsible Use

This model was trained on a mixture of public datasets and synthetic data.
It does not contain personal data by design.
Outputs should not be treated as authoritative in medical, legal, or safety-critical contexts.

Citation & Acknowledgements

If you use this model in research or applications, please acknowledge:

GLM-4.7 for teacher-generated distillation data
Liquid AI for the LFM2 architecture
The creators of the public instruction datasets listed above

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer, TextStreamer

model_id = "yasserrmd/GLM4.7-Distill-LFM2.5-1.2B"

model = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map="auto",
    dtype="bfloat16",
    # attn_implementation="flash_attention_2"  # uncomment on compatible GPU
)

tokenizer = AutoTokenizer.from_pretrained(model_id)
streamer = TextStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)

prompt = "Implement QuickSort algorithm with complexity analysis"

input_ids = tokenizer.apply_chat_template(
    [{"role": "user", "content": prompt}],
    add_generation_prompt=True,
    return_tensors="pt",
    tokenize=True,
).to(model.device)

output = model.generate(
    input_ids,
    do_sample=True,
    temperature=0.1,
    top_k=50,
    top_p=0.1,
    repetition_penalty=1.05,
    max_new_tokens=512,
    streamer=streamer,
)

With vLLM (Production)

python

from vllm import LLM, SamplingParams

llm = LLM(model="yasserrmd/GLM4.7-Distill-LFM2.5-1.2B")

sampling_params = SamplingParams(
    temperature=0.1,
    top_k=50,
    top_p=0.1,
    repetition_penalty=1.05,
    max_tokens=512
)

prompts = [
    "Implement QuickSort algorithm",
    "Solve the Longest Common Subsequence problem",
    "Design a hash table with collision handling"
]

outputs = llm.generate(prompts, sampling_params)

for output in outputs:
    print(output.outputs[0].text)

Recommended Use

Technical interviews
Algorithm learning
Code generation
Problem-solving
Code refactoring
Educational tutoring

Not Recommended For

Current events or recent information
Factual knowledge queries
Legal, medical, or safety-critical code
Highly specialized domain problems
Real-time critical systems without human review

License

Please refer to the licenses of:

the base LFM2 model
the individual datasets used for training

This repository follows the same usage constraints as the upstream components.