Gadsdencode

Nomadic-ICDU-v8

Model details

Table with columns: Field, Value
Field	Value
Model	`Gadsdencode/Nomadic-ICDU-v8`
Model type	Decoder-only causal language model
Architecture	`MistralForCausalLM`
Parameters	7,248,023,552
Base model	`mistralai/Mistral-7B-Instruct-v0.3`
Layers / hidden size	32 / 4,096
Attention heads / KV heads	32 / 8
Vocabulary	32,768 tokens
Declared maximum positions	32,768 tokens
Published precision	FP16
Published local formats	GGUF F16 and GGUF Q4_K_M
Primary language	English
Repository license declaration	MIT
Public checkpoint commit	`a328cf13db7a10a200cf616c2f35a0035a6b637c`
Repository last updated	2025-08-28

Users are responsible for reviewing the upstream base-model terms, the repository license, and any obligations that apply to their deployment or data.

What ICDU-v8 was trained to do

ICDU-v8 is intended to improve task-specific behavior by making the operating contract of a workflow explicit. Depending on the ICDU used for a task, that contract can include:

the user's or organization's intent;
governing principles and priorities;
role, audience, tone, and communication requirements;
domain context and input structure;
constraints, boundaries, and prohibited actions;
decision, escalation, and uncertainty-handling rules; and
examples of preferred and non-preferred behavior.

The intended result is a model that follows a defined operating envelope more consistently than a generic instruction model, including when a request is paraphrased, partially specified, or placed under conflicting pressure.

This is an intended capability, not a guarantee. Users should measure it on their own ICDUs, inputs, failure modes, and deployment conditions.

Base model and training method

The checkpoint configuration identifies mistralai/Mistral-7B-Instruct-v0.3 as the base model. The published repository is a merged inference checkpoint, not a standalone adapter.

ICDU project documentation describes the following training and release methodology:

Convert the target workflow into structured ICDU training records.
Perform supervised fine-tuning with QLoRA.
Apply preference tuning, including DPO, where preferred and rejected responses are available.
Stress-test behavior with perturbations to role, tone, constraints, inputs, and channel.
Evaluate against task-specific gates using an AI Judge and, where appropriate, human-in-the-loop grading.
Release only after the model satisfies the selected operating thresholds.

The public v8 repository does not currently include the run configuration needed to prove which of these stages, hyperparameters, adapter settings, random seeds, checkpoints, or release gates were used for this exact build. Accordingly, the sequence above documents the ICDU project method rather than a reproducible v8 training log.

Training data

ICDU training records are designed to encode task intent and expected behavior rather than rely only on broad, unstructured instruction examples. A record may contain:

an intent or objective;
principles and prioritization rules;
persona, audience, tone, or channel requirements;
relevant context and constraints;
task inputs;
an expected or preferred response;
a rejected response or failure example for preference tuning; and
escalation or abstention behavior.

Data disclosure and exclusions

The following v8-specific information is not present in the public repository:

Table with columns: Disclosure item, Public status
Disclosure item	Public status
Number of training, validation, and test examples	Not published
Dataset names and source provenance	Not published
Human-authored versus synthetic-data composition	Not published
Domain and language distribution	Not published
Deduplication and contamination checks	Not published
Copyright and license review	Not published
PII or sensitive-data screening	Not published
Explicit exclusion list	Not published

No specific exclusion—such as personal data, customer data, copyrighted material, medical records, security-sensitive data, or benchmark test sets—should be assumed without a training-data manifest. Deployments that require documented provenance, consent, data residency, or regulated-data controls should not rely on this checkpoint until the relevant records have been reviewed.

Required prompt and chat format

Use the tokenizer's bundled Mistral chat template whenever possible. Messages must alternate between user and assistant; an optional system message may appear first.

Transformers chat template

python
from transformers import AutoTokenizer

model_id = "Gadsdencode/Nomadic-ICDU-v8"
tokenizer = AutoTokenizer.from_pretrained(model_id)

messages = [
    {
        "role": "system",
        "content": (
            "Follow the supplied ICDU. Respect its intent, principles, "
            "constraints, and escalation rules."
        ),
    },
    {
        "role": "user",
        "content": "Summarize the case and identify any required escalation.",
    },
]

prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)
print(prompt)

Raw single-turn format

text
<s>[INST] {optional system instruction}

{user message} [/INST]

Raw multi-turn format

text
<s>[INST] {system instruction}

{first user message} [/INST] {first assistant response}</s>[INST] {next user message} [/INST]

Do not substitute an unrelated chat template. Prompt-template mismatches can cause lower-quality responses, leaked control tokens, or unstable turn-taking.

The tokenizer also contains formatting for tool calls and tool results. The presence of those tags does not establish reliable tool selection, argument generation, or safe autonomous tool use; evaluate tool behavior separately before enabling it.

Intended uses

Appropriate uses include:

research and evaluation of ICDU-based task specialization;
local or private task-specific assistants;
workflow prototypes with explicit intent, constraints, and escalation rules;
drafting, classification, summarization, and decision support within a tested operating envelope;
policy-sensitive or regulated-workflow support with qualified human review;
perturbation testing and comparison against a base instruction model; and
creating organization- or domain-specific derivatives using authorized data.

The model is best treated as a component inside a governed workflow, not as an independent authority.

Prohibited uses

Do not use ICDU-v8 for:

autonomous medical, legal, financial, employment, insurance, credit, housing, or other high-impact decisions;
emergency response or safety-critical control without qualified human authority;
unsupervised actions that can spend money, modify production systems, disclose data, or create irreversible effects;
unlawful surveillance, discrimination, impersonation, fraud, deception, or harassment;
generation or operational assistance for weapons, malware, credential theft, or other harmful activity;
processing secrets, credentials, personal data, or regulated records without appropriate authorization and technical controls; or
representing model output as guaranteed accurate, compliant, unbiased, or professionally approved.

These use restrictions state the maintainer's intended operating policy. They do not replace applicable law, professional obligations, or license terms.

Known limitations

No published v8 benchmark report. Task-fit and safety claims have not yet been substantiated by reproducible public results.
Incomplete training-data disclosure. Dataset scale, sources, composition, contamination testing, and exclusions are not currently documented.
Hallucinations. The model can produce fluent but false, unsupported, or internally inconsistent content.
Prompt sensitivity. Small changes in instructions, chat formatting, context order, or generation settings can change behavior.
Base-model limits. This is a 7B-class model and should not be expected to match larger frontier systems on broad knowledge, complex reasoning, coding, or multilingual tasks.
Language coverage. English is the declared primary language. Quality in other languages has not been documented.
Long-context behavior is unverified. The configuration declares 32,768 positions, but effective recall and instruction adherence at that length have not been publicly measured for v8.
Tool use is unverified. Tool-format support in the tokenizer is not proof of correct or safe function calling.
Quantization effects. GGUF Q4_K_M is smaller and easier to run locally, but may differ from the FP16 checkpoint in accuracy, formatting, and consistency.
Authentication, authorization, redaction, retrieval permissions, audit logging, and action approval must be provided by the surrounding application.

Benchmark results

No verified, reproducible benchmark scores for Nomadic-ICDU-v8 are currently published. A blank or estimated score would be misleading, so this release does not claim one.

The recommended ICDU evaluation suite is:

Table with columns: Evaluation, What it measures, v8 public result
Evaluation	What it measures	v8 public result
Intent alignment	Completion of the task's explicit objective	Not published
Principle adherence	Compliance with stated priorities and rules	Not published
Application / groundedness	Correct use of supplied inputs and context	Not published
Constraint adherence	Compliance with required and prohibited behavior	Not published
Escalation accuracy	Correct abstention or escalation when a boundary is reached	Not published

For a credible release result, publish the evaluation dataset or a representative, non-sensitive sample; scoring rubric; judge prompts; human-review procedure; sample sizes; confidence intervals; base-model comparison; inference settings; and exact checkpoint hash.

Recommended generation settings

Start with deterministic decoding for governed or repeatable workflows:

Table with columns: Setting, Governed/default, Exploratory drafting
Setting	Governed/default	Exploratory drafting
`do_sample`	`false`	`true`
`temperature`	Omit when sampling is disabled	`0.4`–`0.7`
`top_p`	Omit when sampling is disabled	–

Example:

python
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Gadsdencode/Nomadic-ICDU-v8"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
)

messages = [
    {"role": "system", "content": "Follow the supplied ICDU and its escalation rules."},
    {"role": "user", "content": "Review this input and return the required action."},
]
inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    return_tensors="pt",
).to(model.device)

outputs = model.generate(
    inputs,
    do_sample=False,
    max_new_tokens=512,
    repetition_penalty=1.05,
    eos_token_id=tokenizer.eos_token_id,
)

new_tokens = outputs[0, inputs.shape[-1]:]
print(tokenizer.decode(new_tokens, skip_special_tokens=True))

Generation settings should be selected against the target evaluation suite. Lower temperature improves repeatability but does not guarantee correctness or policy compliance.

Context-length guidance

The model configuration declares a 32,768-token maximum position length. That is an architectural limit, not a promise that all 32K-token prompts will be recalled or followed equally well.

Table with columns: Use case, Recommended starting context
Use case	Recommended starting context
Local Q4 testing	8,192 tokens
Typical hosted workflow	8,192–16,384 tokens
Long-document evaluation	16,384–32,768 tokens
Production at 32K	Only after task-specific recall and adherence tests

Increasing context length raises KV-cache memory use, latency, and cost and can reduce concurrency. Prefer retrieving the smallest relevant context, placing the ICDU and critical constraints clearly, and testing required facts near the beginning, middle, and end of long prompts.

If a local application reports an 8,192-token limit, that is usually its runtime configuration rather than a different model architecture. Increase the runtime setting only if the available RAM or VRAM and evaluation results support it.

Local use with LM Studio

In LM Studio, search for Gadsdencode/Nomadic-ICDU-v8.
Download nomadic-icdu-v8-Q4_K_M.gguf for the practical local build. Use the F16 GGUF only when the additional memory requirement is acceptable.
Load the model and leave the prompt template on Auto or select the Mistral Instruct template.
Start with an 8,192-token context. Move to 16,384 or 32,768 only after checking memory use and long-context behavior.
For deterministic workflows, disable sampling or set temperature to the lowest supported value.
To expose a local API, start LM Studio's OpenAI-compatible server. The default is commonly http://127.0.0.1:1234/v1.

Check the model identifier returned by GET /v1/models, then use it in the request:

bash
curl http://127.0.0.1:1234/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "replace-with-the-id-from-v1-models",
    "messages": [
      {
        "role": "system",
        "content": "Follow the supplied ICDU and its escalation rules."
      },
      {
        "role": "user",
        "content": "Review this input and return the required action."
      }
    ],
    "temperature": 0,
    "max_tokens": 512
  }'

Local use with Ollama

Download the Q4_K_M GGUF:

bash
hf download Gadsdencode/Nomadic-ICDU-v8 \
  nomadic-icdu-v8-Q4_K_M.gguf \
  --local-dir ./nomadic-icdu-v8

Create nomadic-icdu-v8/Modelfile:

text
FROM ./nomadic-icdu-v8-Q4_K_M.gguf

PARAMETER num_ctx 8192
PARAMETER num_predict 512
PARAMETER temperature 0
PARAMETER top_p 0.9
PARAMETER repeat_penalty 1.05
PARAMETER stop "</s>"

Create and run the local model:

bash
cd nomadic-icdu-v8
ollama create nomadic-icdu-v8 -f Modelfile
ollama run nomadic-icdu-v8

Ollama should read the chat-template metadata embedded in the GGUF. Confirm during acceptance testing that prompts render with Mistral [INST] ... [/INST] formatting.

Ollama API example:

bash
curl http://localhost:11434/api/chat \
  -H "Content-Type: application/json" \
  -d '{
    "model": "nomadic-icdu-v8",
    "stream": false,
    "messages": [
      {
        "role": "system",
        "content": "Follow the supplied ICDU and its escalation rules."
      },
      {
        "role": "user",
        "content": "Review this input and return the required action."
      }
    ],
    "options": {
      "temperature": 0,
      "num_ctx": 8192,
      "num_predict": 512
    }
  }'

Hosted API example

For a hosted deployment, use the FP16 Safetensors checkpoint with an inference engine that honors the tokenizer chat template. An OpenAI-compatible API makes it easier to move between a hosted endpoint, LM Studio, and other serving systems.

Set:

bash
export ICDU_API_BASE_URL="https://your-endpoint.example.com/v1"
export ICDU_API_KEY="replace-with-your-endpoint-token"

Python:

python
import os
from openai import OpenAI

client = OpenAI(
    base_url=os.environ["ICDU_API_BASE_URL"],
    api_key=os.environ["ICDU_API_KEY"],
)

response = client.chat.completions.create(
    model="Gadsdencode/Nomadic-ICDU-v8",
    messages=[
        {
            "role": "system",
            "content": "Follow the supplied ICDU and its escalation rules.",
        },
        {
            "role": "user",
            "content": "Review this input and return the required action.",
        },
    ],
    temperature=0,
    max_tokens=512,
)

print(response.choices[0].message.content)

Endpoint operators should add authentication, request-size limits, timeouts, structured logging with sensitive-data controls, rate limits, abuse monitoring, and human approval around consequential actions. Do not log raw prompts or completions by default when they may contain confidential data.

Model and version lineage

text
mistralai/Mistral-7B-Instruct-v0.3
└── Nomadic-ICDU-v8 merged FP16 checkpoint
    ├── Transformers Safetensors (3 shards)
    ├── GGUF F16
    └── GGUF Q4_K_M

v8 is the project's release label for this checkpoint. Public artifacts describing v1 through v7, their training changes, and their comparative evaluations are not included in this repository, so no undocumented lineage claims are made here.

For reproducibility, downstream derivatives should record:

the exact source revision or checkpoint hash;
the ICDU and dataset manifest version;
training code and dependency revisions;
training and preference-tuning hyperparameters;
random seeds;
evaluation-suite version and release thresholds;
quantization tool, version, and parameters; and
any changes to the prompt template or context configuration.

Citation

If you use this model, cite the model repository and the exact revision:

bibtex
@misc{nomadic_icdu_v8,
  title        = {Nomadic-ICDU-v8},
  author       = {{Gadsdencode}},
  year         = {2025},
  howpublished = {\url{https://huggingface.co/Gadsdencode/Nomadic-ICDU-v8}},
  note         = {Revision a328cf13db7a10a200cf616c2f35a0035a6b637c}
}

Additional information

Model repository: https://huggingface.co/Gadsdencode/Nomadic-ICDU-v8
Base model: https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.3

The bundled handler.py should be reviewed before production use. In particular, verify that it loads the correct model ID, applies the tokenizer's chat template, does not expose exception details to end users, and does not print sensitive prompts or completions.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Container

Run this model inference with full control and performance in your environment.

Learn more

Model Details

Model Provider

Gadsdencode

Model Tree

Base

mistralai/Mistral-7B-Instruct-v0.3

Fine-tuned

this model

Input Modalities

Text

Output Modalities

Text

Supported Functionality

Dedicated EndpointsContainer

Explore FriendliAI today

Get started Talk to an engineer

Model details

Table with columns: Field, Value
Field	Value
Model	`Gadsdencode/Nomadic-ICDU-v8`
Model type	Decoder-only causal language model
Architecture	`MistralForCausalLM`
Parameters	7,248,023,552
Base model	`mistralai/Mistral-7B-Instruct-v0.3`
Layers / hidden size	32 / 4,096
Attention heads / KV heads	32 / 8
Vocabulary	32,768 tokens
Declared maximum positions	32,768 tokens
Published precision	FP16
Published local formats	GGUF F16 and GGUF Q4_K_M
Primary language	English
Repository license declaration	MIT
Public checkpoint commit	`a328cf13db7a10a200cf616c2f35a0035a6b637c`
Repository last updated	2025-08-28

Users are responsible for reviewing the upstream base-model terms, the repository license, and any obligations that apply to their deployment or data.

What ICDU-v8 was trained to do

ICDU-v8 is intended to improve task-specific behavior by making the operating contract of a workflow explicit. Depending on the ICDU used for a task, that contract can include:

the user's or organization's intent;
governing principles and priorities;
role, audience, tone, and communication requirements;
domain context and input structure;
constraints, boundaries, and prohibited actions;
decision, escalation, and uncertainty-handling rules; and
examples of preferred and non-preferred behavior.

This is an intended capability, not a guarantee. Users should measure it on their own ICDUs, inputs, failure modes, and deployment conditions.

Base model and training method

The checkpoint configuration identifies mistralai/Mistral-7B-Instruct-v0.3 as the base model. The published repository is a merged inference checkpoint, not a standalone adapter.

ICDU project documentation describes the following training and release methodology:

Convert the target workflow into structured ICDU training records.
Perform supervised fine-tuning with QLoRA.
Apply preference tuning, including DPO, where preferred and rejected responses are available.
Stress-test behavior with perturbations to role, tone, constraints, inputs, and channel.
Evaluate against task-specific gates using an AI Judge and, where appropriate, human-in-the-loop grading.
Release only after the model satisfies the selected operating thresholds.

Training data

ICDU training records are designed to encode task intent and expected behavior rather than rely only on broad, unstructured instruction examples. A record may contain:

an intent or objective;
principles and prioritization rules;
persona, audience, tone, or channel requirements;
relevant context and constraints;
task inputs;
an expected or preferred response;
a rejected response or failure example for preference tuning; and
escalation or abstention behavior.

Data disclosure and exclusions

The following v8-specific information is not present in the public repository:

Table with columns: Disclosure item, Public status
Disclosure item	Public status
Number of training, validation, and test examples	Not published
Dataset names and source provenance	Not published
Human-authored versus synthetic-data composition	Not published
Domain and language distribution	Not published
Deduplication and contamination checks	Not published
Copyright and license review	Not published
PII or sensitive-data screening	Not published
Explicit exclusion list	Not published

Required prompt and chat format

Use the tokenizer's bundled Mistral chat template whenever possible. Messages must alternate between user and assistant; an optional system message may appear first.

Transformers chat template

python
from transformers import AutoTokenizer

model_id = "Gadsdencode/Nomadic-ICDU-v8"
tokenizer = AutoTokenizer.from_pretrained(model_id)

messages = [
    {
        "role": "system",
        "content": (
            "Follow the supplied ICDU. Respect its intent, principles, "
            "constraints, and escalation rules."
        ),
    },
    {
        "role": "user",
        "content": "Summarize the case and identify any required escalation.",
    },
]

prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)
print(prompt)

Raw single-turn format

text
<s>[INST] {optional system instruction}

{user message} [/INST]

Raw multi-turn format

text
<s>[INST] {system instruction}

{first user message} [/INST] {first assistant response}</s>[INST] {next user message} [/INST]

Do not substitute an unrelated chat template. Prompt-template mismatches can cause lower-quality responses, leaked control tokens, or unstable turn-taking.

Intended uses

Appropriate uses include:

research and evaluation of ICDU-based task specialization;
local or private task-specific assistants;
workflow prototypes with explicit intent, constraints, and escalation rules;
drafting, classification, summarization, and decision support within a tested operating envelope;
policy-sensitive or regulated-workflow support with qualified human review;
perturbation testing and comparison against a base instruction model; and
creating organization- or domain-specific derivatives using authorized data.

The model is best treated as a component inside a governed workflow, not as an independent authority.

Prohibited uses

Do not use ICDU-v8 for:

autonomous medical, legal, financial, employment, insurance, credit, housing, or other high-impact decisions;
emergency response or safety-critical control without qualified human authority;
unsupervised actions that can spend money, modify production systems, disclose data, or create irreversible effects;
unlawful surveillance, discrimination, impersonation, fraud, deception, or harassment;
generation or operational assistance for weapons, malware, credential theft, or other harmful activity;
processing secrets, credentials, personal data, or regulated records without appropriate authorization and technical controls; or
representing model output as guaranteed accurate, compliant, unbiased, or professionally approved.

These use restrictions state the maintainer's intended operating policy. They do not replace applicable law, professional obligations, or license terms.

Known limitations

No published v8 benchmark report. Task-fit and safety claims have not yet been substantiated by reproducible public results.
Incomplete training-data disclosure. Dataset scale, sources, composition, contamination testing, and exclusions are not currently documented.
Hallucinations. The model can produce fluent but false, unsupported, or internally inconsistent content.
Prompt sensitivity. Small changes in instructions, chat formatting, context order, or generation settings can change behavior.
Base-model limits. This is a 7B-class model and should not be expected to match larger frontier systems on broad knowledge, complex reasoning, coding, or multilingual tasks.
Language coverage. English is the declared primary language. Quality in other languages has not been documented.
Long-context behavior is unverified. The configuration declares 32,768 positions, but effective recall and instruction adherence at that length have not been publicly measured for v8.
Tool use is unverified. Tool-format support in the tokenizer is not proof of correct or safe function calling.
Quantization effects. GGUF Q4_K_M is smaller and easier to run locally, but may differ from the FP16 checkpoint in accuracy, formatting, and consistency.
Authentication, authorization, redaction, retrieval permissions, audit logging, and action approval must be provided by the surrounding application.

Benchmark results

No verified, reproducible benchmark scores for Nomadic-ICDU-v8 are currently published. A blank or estimated score would be misleading, so this release does not claim one.

The recommended ICDU evaluation suite is:

Table with columns: Evaluation, What it measures, v8 public result
Evaluation	What it measures	v8 public result
Intent alignment	Completion of the task's explicit objective	Not published
Principle adherence	Compliance with stated priorities and rules	Not published
Application / groundedness	Correct use of supplied inputs and context	Not published
Constraint adherence	Compliance with required and prohibited behavior	Not published
Escalation accuracy	Correct abstention or escalation when a boundary is reached	Not published

Recommended generation settings

Start with deterministic decoding for governed or repeatable workflows:

Table with columns: Setting, Governed/default, Exploratory drafting
Setting	Governed/default	Exploratory drafting
`do_sample`	`false`	`true`
`temperature`	Omit when sampling is disabled	`0.4`–`0.7`
`top_p`	Omit when sampling is disabled	–

Example:

python
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Gadsdencode/Nomadic-ICDU-v8"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
)

messages = [
    {"role": "system", "content": "Follow the supplied ICDU and its escalation rules."},
    {"role": "user", "content": "Review this input and return the required action."},
]
inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    return_tensors="pt",
).to(model.device)

outputs = model.generate(
    inputs,
    do_sample=False,
    max_new_tokens=512,
    repetition_penalty=1.05,
    eos_token_id=tokenizer.eos_token_id,
)

new_tokens = outputs[0, inputs.shape[-1]:]
print(tokenizer.decode(new_tokens, skip_special_tokens=True))

Generation settings should be selected against the target evaluation suite. Lower temperature improves repeatability but does not guarantee correctness or policy compliance.

Context-length guidance

The model configuration declares a 32,768-token maximum position length. That is an architectural limit, not a promise that all 32K-token prompts will be recalled or followed equally well.

Table with columns: Use case, Recommended starting context
Use case	Recommended starting context
Local Q4 testing	8,192 tokens
Typical hosted workflow	8,192–16,384 tokens
Long-document evaluation	16,384–32,768 tokens
Production at 32K	Only after task-specific recall and adherence tests

Local use with LM Studio

In LM Studio, search for Gadsdencode/Nomadic-ICDU-v8.
Download nomadic-icdu-v8-Q4_K_M.gguf for the practical local build. Use the F16 GGUF only when the additional memory requirement is acceptable.
Load the model and leave the prompt template on Auto or select the Mistral Instruct template.
Start with an 8,192-token context. Move to 16,384 or 32,768 only after checking memory use and long-context behavior.
For deterministic workflows, disable sampling or set temperature to the lowest supported value.
To expose a local API, start LM Studio's OpenAI-compatible server. The default is commonly http://127.0.0.1:1234/v1.

Check the model identifier returned by GET /v1/models, then use it in the request:

bash
curl http://127.0.0.1:1234/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "replace-with-the-id-from-v1-models",
    "messages": [
      {
        "role": "system",
        "content": "Follow the supplied ICDU and its escalation rules."
      },
      {
        "role": "user",
        "content": "Review this input and return the required action."
      }
    ],
    "temperature": 0,
    "max_tokens": 512
  }'

Local use with Ollama

Download the Q4_K_M GGUF:

bash
hf download Gadsdencode/Nomadic-ICDU-v8 \
  nomadic-icdu-v8-Q4_K_M.gguf \
  --local-dir ./nomadic-icdu-v8

Create nomadic-icdu-v8/Modelfile:

text
FROM ./nomadic-icdu-v8-Q4_K_M.gguf

PARAMETER num_ctx 8192
PARAMETER num_predict 512
PARAMETER temperature 0
PARAMETER top_p 0.9
PARAMETER repeat_penalty 1.05
PARAMETER stop "</s>"

Create and run the local model:

bash
cd nomadic-icdu-v8
ollama create nomadic-icdu-v8 -f Modelfile
ollama run nomadic-icdu-v8

Ollama should read the chat-template metadata embedded in the GGUF. Confirm during acceptance testing that prompts render with Mistral [INST] ... [/INST] formatting.

Ollama API example:

bash
curl http://localhost:11434/api/chat \
  -H "Content-Type: application/json" \
  -d '{
    "model": "nomadic-icdu-v8",
    "stream": false,
    "messages": [
      {
        "role": "system",
        "content": "Follow the supplied ICDU and its escalation rules."
      },
      {
        "role": "user",
        "content": "Review this input and return the required action."
      }
    ],
    "options": {
      "temperature": 0,
      "num_ctx": 8192,
      "num_predict": 512
    }
  }'

Hosted API example

Set:

bash
export ICDU_API_BASE_URL="https://your-endpoint.example.com/v1"
export ICDU_API_KEY="replace-with-your-endpoint-token"

Python:

python
import os
from openai import OpenAI

client = OpenAI(
    base_url=os.environ["ICDU_API_BASE_URL"],
    api_key=os.environ["ICDU_API_KEY"],
)

response = client.chat.completions.create(
    model="Gadsdencode/Nomadic-ICDU-v8",
    messages=[
        {
            "role": "system",
            "content": "Follow the supplied ICDU and its escalation rules.",
        },
        {
            "role": "user",
            "content": "Review this input and return the required action.",
        },
    ],
    temperature=0,
    max_tokens=512,
)

print(response.choices[0].message.content)

Model and version lineage

text
mistralai/Mistral-7B-Instruct-v0.3
└── Nomadic-ICDU-v8 merged FP16 checkpoint
    ├── Transformers Safetensors (3 shards)
    ├── GGUF F16
    └── GGUF Q4_K_M

For reproducibility, downstream derivatives should record:

the exact source revision or checkpoint hash;
the ICDU and dataset manifest version;
training code and dependency revisions;
training and preference-tuning hyperparameters;
random seeds;
evaluation-suite version and release thresholds;
quantization tool, version, and parameters; and
any changes to the prompt template or context configuration.

Citation

If you use this model, cite the model repository and the exact revision:

bibtex
@misc{nomadic_icdu_v8,
  title        = {Nomadic-ICDU-v8},
  author       = {{Gadsdencode}},
  year         = {2025},
  howpublished = {\url{https://huggingface.co/Gadsdencode/Nomadic-ICDU-v8}},
  note         = {Revision a328cf13db7a10a200cf616c2f35a0035a6b637c}
}

Additional information

Model repository: https://huggingface.co/Gadsdencode/Nomadic-ICDU-v8
Base model: https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.3

Nomadic-ICDU-v8

README

Model details

What ICDU-v8 was trained to do

Base model and training method

Training data

Data disclosure and exclusions

Required prompt and chat format

Transformers chat template

Raw single-turn format

Raw multi-turn format

Intended uses

Prohibited uses

Known limitations

Benchmark results

Recommended generation settings

Context-length guidance

Local use with LM Studio

Local use with Ollama

Hosted API example

Model and version lineage

Citation

Additional information

Explore FriendliAI today

README

Model details

What ICDU-v8 was trained to do

Base model and training method

Training data

Data disclosure and exclusions

Required prompt and chat format

Transformers chat template

Raw single-turn format

Raw multi-turn format

Intended uses

Prohibited uses

Known limitations

Benchmark results

Recommended generation settings

Context-length guidance

Local use with LM Studio

Local use with Ollama

Hosted API example

Model and version lineage

Citation

Additional information