Pomoika24

C1-Tachu

README

License: apache-2.0

Model Details

Table

Base model	Qwen2.5-Coder-14B
Architecture	Qwen2 (Causal LM)
Parameters	14B
Hidden size	5120
Layers	48
Attention heads	40 (8 KV heads, GQA)
Context length	32,768 tokens
Precision	bfloat16
Format	Qwen ChatML
License	Apache 2.0

Training

C1-Tachu was trained using QLoRA (4-bit nf4 quantization with LoRA adapters) via the Axolotl framework on AWS g5.2xlarge (NVIDIA A10G).

Pipeline (8 stages)

Each stage trains a fresh LoRA adapter on the previous stage's merged model, then merges it back into the base weights. Adapters are not stacked — each stage builds cleanly on the merged result.

Table with columns: Stage, Name, LoRA Rank, Examples, Loss
Stage	Name	LoRA Rank	Examples	Loss
1	Software Foundation	64	~10,000	—
2	Fast Code Generation	64	~6,000	—
3	Debugging Speed	64	~7,000	—
4

Stage 7 (ORPO) was skipped in this training run due to time constraints. The model retains all capabilities from stages 1–6 and 8.

QLoRA Hyperparameters

Table with columns: Parameter, Value
Parameter	Value
Quantization	nf4, double quantization
Compute dtype	bfloat16
LoRA alpha	128 (2× rank)
LoRA dropout	0.05
LoRA target modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Optimizer	paged_adamw_8bit
Learning rate	1e-4
LR scheduler	cosine
Warmup ratio

Capabilities

C1-Tachu is trained across six skill domains:

Software Foundation — Broad code familiarity across languages and patterns, calibrated to a terse, correct response style
Fast Code Generation — Produces correct implementations with minimal tokens; no unnecessary comments or padding
Debugging Speed — Jumps to root cause instead of enumerating possibilities; trained on real GitHub commit-fix pairs
Project Navigation — Answers "where" and "how" questions about unfamiliar repos without reading every file
Rapid Architecture — Compresses the planning→coding loop: requirements → architecture → plan → code skeleton
Self Verification — Agent loop with tool use (run_tests, compile, read_file, write_file, run_command) to reach working solutions with minimal human intervention
Agent Workflows — Complex multi-step task decomposition with parallel tool calls

Quickstart

python
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "Pomoika24/C1-Tachu"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto",
)

messages = [
    {"role": "system", "content": "You are Tachu, a fast software engineer."},
    {"role": "user", "content": "Implement a retry wrapper with exponential backoff in Python."},
]

text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=2048,
)
response = tokenizer.decode(generated_ids[0][len(model_inputs.input_ids[0]):], skip_special_tokens=True)
print(response)

Deployment with vLLM

bash
pip install vllm
vllm serve "Pomoika24/C1-Tachu"

Deployment with Ollama

bash
ollama run hf.co/Pomoika24/C1-Tachu

Security Posture

Stance: passive (do no harm).

Never generates known-vulnerable patterns: hardcoded secrets, SQL injection, command injection, path traversal, XSS, insecure deserialization, disabled auth checks
Defaults to safe patterns: parameterized queries, input validation, proper error handling without leaking stack traces
Does not actively scan code for vulnerabilities unless explicitly asked for a security review
Does not refuse to work on codebases with existing vulnerabilities — just doesn't make them worse
When uncertain, prefers the safer pattern even if more verbose

Limitations

Stage 7 (ORPO) not trained — the model has not undergone preference optimization, so it may produce verbose solutions where a shorter one would suffice
Training data not published — the dataset was generated on cloud infrastructure and is not publicly available
Single-annotator evaluation — eval sets were scored by one developer with LLM-as-judge assistance; no multi-annotator agreement metrics
Context limit — while the base model supports 32K context, training stages used 2048–4096 token sequences; performance may degrade on very long contexts
English-focused — training data is predominantly English

Intended Use

C1-Tachu is designed for:

Software engineers looking to accelerate their workflow
Coding agents and assistants that need fast, correct code generation
Automated debugging and code navigation tasks
Architecture planning and code scaffolding

Not intended for:

Automated security auditing (passive stance only)
Code generation in safety-critical systems without human review
Production deployment without human oversight

Citation

bibtex
@misc{c1-tachu,
  title={C1-Tachu: A Speed-Optimized Coding AI},
  author={Vladislav Kondratyev},
  year={2026},
  url={https://huggingface.co/Pomoika24/C1-Tachu}
}

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

Pomoika24

Model Tree

Base

Qwen/Qwen2.5-Coder-14B

Fine-tuned

this model

Input Modalities

Text

Output Modalities

Text

Supported Functionality

Dedicated EndpointsContainer

Explore FriendliAI today

Get started Talk to an engineer

README

License: apache-2.0

Model Details

Table

Base model	Qwen2.5-Coder-14B
Architecture	Qwen2 (Causal LM)
Parameters	14B
Hidden size	5120
Layers	48
Attention heads	40 (8 KV heads, GQA)
Context length	32,768 tokens
Precision	bfloat16
Format	Qwen ChatML
License	Apache 2.0

Training

C1-Tachu was trained using QLoRA (4-bit nf4 quantization with LoRA adapters) via the Axolotl framework on AWS g5.2xlarge (NVIDIA A10G).

Pipeline (8 stages)

Each stage trains a fresh LoRA adapter on the previous stage's merged model, then merges it back into the base weights. Adapters are not stacked — each stage builds cleanly on the merged result.

Table with columns: Stage, Name, LoRA Rank, Examples, Loss
Stage	Name	LoRA Rank	Examples	Loss
1	Software Foundation	64	~10,000	—
2	Fast Code Generation	64	~6,000	—
3	Debugging Speed	64	~7,000	—
4

Stage 7 (ORPO) was skipped in this training run due to time constraints. The model retains all capabilities from stages 1–6 and 8.

QLoRA Hyperparameters

Table with columns: Parameter, Value
Parameter	Value
Quantization	nf4, double quantization
Compute dtype	bfloat16
LoRA alpha	128 (2× rank)
LoRA dropout	0.05
LoRA target modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Optimizer	paged_adamw_8bit
Learning rate	1e-4
LR scheduler	cosine
Warmup ratio

Capabilities

C1-Tachu is trained across six skill domains:

Software Foundation — Broad code familiarity across languages and patterns, calibrated to a terse, correct response style
Fast Code Generation — Produces correct implementations with minimal tokens; no unnecessary comments or padding
Debugging Speed — Jumps to root cause instead of enumerating possibilities; trained on real GitHub commit-fix pairs
Project Navigation — Answers "where" and "how" questions about unfamiliar repos without reading every file
Rapid Architecture — Compresses the planning→coding loop: requirements → architecture → plan → code skeleton
Self Verification — Agent loop with tool use (run_tests, compile, read_file, write_file, run_command) to reach working solutions with minimal human intervention
Agent Workflows — Complex multi-step task decomposition with parallel tool calls

Quickstart

python
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "Pomoika24/C1-Tachu"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto",
)

messages = [
    {"role": "system", "content": "You are Tachu, a fast software engineer."},
    {"role": "user", "content": "Implement a retry wrapper with exponential backoff in Python."},
]

text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=2048,
)
response = tokenizer.decode(generated_ids[0][len(model_inputs.input_ids[0]):], skip_special_tokens=True)
print(response)

Deployment with vLLM

bash
pip install vllm
vllm serve "Pomoika24/C1-Tachu"

Deployment with Ollama

bash
ollama run hf.co/Pomoika24/C1-Tachu

Security Posture

Stance: passive (do no harm).

Never generates known-vulnerable patterns: hardcoded secrets, SQL injection, command injection, path traversal, XSS, insecure deserialization, disabled auth checks
Defaults to safe patterns: parameterized queries, input validation, proper error handling without leaking stack traces
Does not actively scan code for vulnerabilities unless explicitly asked for a security review
Does not refuse to work on codebases with existing vulnerabilities — just doesn't make them worse
When uncertain, prefers the safer pattern even if more verbose

Limitations

Stage 7 (ORPO) not trained — the model has not undergone preference optimization, so it may produce verbose solutions where a shorter one would suffice
Training data not published — the dataset was generated on cloud infrastructure and is not publicly available
Single-annotator evaluation — eval sets were scored by one developer with LLM-as-judge assistance; no multi-annotator agreement metrics
Context limit — while the base model supports 32K context, training stages used 2048–4096 token sequences; performance may degrade on very long contexts
English-focused — training data is predominantly English

Intended Use

C1-Tachu is designed for:

Software engineers looking to accelerate their workflow
Coding agents and assistants that need fast, correct code generation
Automated debugging and code navigation tasks
Architecture planning and code scaffolding

Not intended for:

Automated security auditing (passive stance only)
Code generation in safety-critical systems without human review
Production deployment without human oversight

Citation

bibtex
@misc{c1-tachu,
  title={C1-Tachu: A Speed-Optimized Coding AI},
  author={Vladislav Kondratyev},
  year={2026},
  url={https://huggingface.co/Pomoika24/C1-Tachu}
}