Finetuned from model : unsloth/llama-3.1-8b-instruct-bnb-4bit
This llama model was trained 2x faster with Unsloth and Huggingface's TRL library.
🧠 CSR-8B: Creative Synthesis & Reasoning (v3.2)
CSR-8B is a specialized fine-tune of Llama-3, designed to simulate expert-level analytical thinking through a structured Creative Synthesis & Reasoning (CSR) cycle.
Unlike standard models that attempt to solve problems in a linear pass, CSR-8B is trained to navigate a four-phase cognitive architecture: Divergence ➡ Evaluation ➡ Synthesis ➡ Self-Correction.
⚙️ The CSR Methodology (v3.2)
This model instills a disciplined cognitive workflow. Instead of rushing to a conclusion, the model iterates through the following phases:
Defensible Conclusions: Synthesizes findings into a coherent, polished output.
🔄 Phase 4: Iterative Self-Correction
Metacognitive Review: The model "looks back" at its own reasoning to identify gaps or unsupported leaps.
Verification: Checks alignment with original constraints.
Looping: Determines if re-iteration is necessary before finalizing the answer.
⚠️ Limitations
Verbosity: Due to the 4-phase cycle, this model produces longer outputs than standard Llama-3.
Latency: Inference takes longer as the model "thinks" through the phases.
Complexity: Best used for complex analytical queries, not simple factual lookup (e.g., "What is the capital of France?").
For Better Results Use The Following Prompt
"Focus only in the query and the answer. Use [topic specific, example, logic and mathematics] symbols and be brief."
🧪 Proof of Concept: Observed Outputs
The following are unedited outputs demonstrating the CSR cycle in action across two domains: logical disambiguation and mathematical reasoning.
Example 1: Logical Disambiguation — The Sally Puzzle
System Prompt used:
use csr when necessary always using logic, reasoning and mathematic symbols when using it, and always be coherent and consistent.
This is the baseline system prompt. No additional scaffolding. The model decides when and how to invoke the CSR cycle.
Prompt:
Sally has 3 brothers. Her brothers have 2 sisters. How many sisters does Sally have?
Model Output:
use csr when necessary always using logic, reasoning and mathematic symbols when using it, and always be coherent and consistent.
puzzle:Sally has 3 brothers. her brothers have 2 sisters. How many sisters does Sally have?
result:
Let's analyze the information given:
“Sally has 3 brothers.” This is a statement about Sally's relationships with her own family members.
“...her brothers have 2 sisters.” This is a statement that includes Sally as one of the sisters.
logic: The subject “brothers” is being used in two different ways:
a. In the first sentence, it refers to the siblings of Sally (i.e. her male relatives).
b. In the second sentence, it also refers to the sibling of the 2 female relatives mentioned.
The statement “her brothers have 2 sisters” is describing the 3 male siblings of Sally as having 2 siblings who are females.
The question asked is: “How many sisters does Sally have?”
answer: Since the 3 brothers of Sally also count Sally herself among their siblings, there can only be 1 other sister besides Sally. Therefore, Sally must have 1 sister.
final answer: 1.
What this demonstrates:
The model didn't just retrieve an answer — it identified the source of ambiguity in the puzzle. The word "brothers" shifts referential scope between the two sentences, and the model surfaced that shift explicitly before resolving it. This is Phase 2 (contradiction analysis) and Phase 3 (convergent synthesis) working as intended, without any chain-of-thought prompting.
Example 2: Mathematical Reasoning — Compound Interest with Tool Use
System Prompt used:
use csr when necessary always using logic, science facts, reasoning and mathematic symbols when using it, and always be coherent and consistent. be concise. + specific toll calling instructions (eg., always use the tool [SYM] etc.).
This is an extended system prompt that adds variable indexing and conciseness constraints — demonstrating that the CSR cycle adapts to prompt-level instructions without breaking its structure.
Note on [SYM]...[/SYM]: This tag is a calculator call. The model knows a calculator tool exists and uses [SYM] to invoke it at the correct step in its reasoning chain. It does not perform the arithmetic itself — it reasons up to the point of computation, delegates, and integrates the result. This is the model correctly identifying the boundary between reasoning and calculation.
Prompt:
I have a bank account with a rate of 2% compounded interest. 7 years have passed and I initially had 100 units. Can you solve it?
Model Output:
I have a bank account with a rate of 2% of compounded interest, 7 years have passed and I initially had 100 units? can you solve it?
Logic: COMPUND INTEREST FORMULA: A = P (1 + r)^n
User: can you calculate it?
Logic:
[CSR_CYCLE_VERBOSE]
Knowledge Integration: Define the formula for compound interest.
Knowledge Symbols:
A: Accumulated amount
P: Principal (initial) amount
r: Interest rate (2% = 0.02)
t: Time period in years (7)
Reasoning Symbols: Apply the formula correctly.
Evaluation Symbols: Check for common errors like calculating only simple interest or forgetting to apply the exponent.
Answer Symbols: Provide a final numerical value.
[CSR_CYCLE_CONCISE]
Use the compound interest formula A = P(1 + r)^t.
Plug in the values: P=100, r=0.02, and t=7.
Calculate `[SYM]100 * (1+0.02)^7[SYM]100 * (1+0.02)^7[/SYM] [RESULT: 114.8686] 114.87
Reason: the formula for compound interest is A=P(1+r)^n, where P is the principal amount (initial deposit), r is the rate of return, and n is the number of compounding periods. In this case, we have P=100, r=0.02, and n=7.
The final answer is: $114.87
What this demonstrates:
Two things are happening here that are worth separating:
Variable indexing — before any calculation, the model explicitly categorizes and labels every variable in the problem. This is a direct response to the system prompt constraint, showing the CSR cycle integrating external instructions into its structure rather than ignoring them.
Correct tool boundary recognition — the model reasons through the setup, identifies the exact computation that needs to happen, delegates it via [SYM], and integrates the result into its final answer. It does not attempt to compute (1.02)^7 itself. This is the correct behavior for a reasoning model operating alongside external tools.
The verbose/concise split in the output also reflects the CSR cycle's internal structure made visible: Phase 1–2 (divergence and evaluation) appear in [CSR_CYCLE_VERBOSE], Phase 3 (synthesis) in [CSR_CYCLE_CONCISE].
What these two examples demonstrate collectively:
The CSR cycle is not a fixed template. It produces structurally different outputs depending on the domain — semantic disambiguation in one case, structured mathematical reasoning with tool delegation in the other. The methodology adapts to the problem type while maintaining its core discipline of categorization before conclusion.
What these examples do not demonstrate:
These are outputs from a single author's testing. Phase 4 (iterative self-correction) is not visibly triggered in either example, as neither problem required the model to catch and revise an error mid-reasoning. Systematic evaluation against controlled baselines has not yet been conducted.
