CIRF-Qwen3-8B-Instruct-Full

A CIRF reasoning model fine-tuned from Qwen3-8B, trained with full explicit result supervision: every intermediate result is written out in the training targets. The LoRA update is merged into the dense BF16 weights and the 514 new tokens ([sof], [eof], [f_0] to [f_511]) are included with their trained embedding and output rows, so the model loads with plain Transformers (no PEFT, base checkpoint or CIRF code needed).

Prompt format

QUESTION[sof]

No chat template. The model writes a trace of functional tokens with optional result values, then [eof] and the final answer, and stops at the tokenizer's EOS token. Stop on EOS, not on [eof], and decode with skip_special_tokens=False to keep the boundaries.

Inference

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

repo_id = "JhCIFT/CIRF-Qwen3-8B-Instruct-Full"
tokenizer = AutoTokenizer.from_pretrained(repo_id)
model = AutoModelForCausalLM.from_pretrained(repo_id, dtype=torch.bfloat16, device_map="auto").eval()

prompt = "What is 2 + 2?[sof]"
inputs = tokenizer(prompt, add_special_tokens=False, return_tensors="pt").to(model.device)
with torch.inference_mode():
    output = model.generate(**inputs, max_new_tokens=384, do_sample=False,
                            eos_token_id=tokenizer.eos_token_id, pad_token_id=tokenizer.pad_token_id)
ids = output[0, inputs.input_ids.shape[1]:].tolist()
if tokenizer.eos_token_id in ids:
    ids = ids[:ids.index(tokenizer.eos_token_id)]
trace = tokenizer.decode(ids, skip_special_tokens=False)
print(trace)
print(trace.split("[eof]", 1)[1].strip() if "[eof]" in trace else None)

The evaluation below additionally limits each trace to 32 functional tokens (forcing [eof] when the limit is reached). The included inference.py reproduces that protocol and needs only Torch and Transformers:

python inference.py --model JhCIFT/CIRF-Qwen3-8B-Instruct-Full --question "What is 2 + 2?"

Test accuracy

Test set	Questions	Accuracy
GSM8K	1,319	73.62%
MultiArith	178	100.00%
SVAMP	300	87.00%
CommonsenseQA	1,221	82.56%
Coin Flip (parity-corrected)	3,333	99.73%

Greedy decoding, at most 384 new tokens and 32 functional tokens, exact-match answer scoring. Coin Flip is scored against the parity of the flips stated in the question, because about 20% of the dataset's labels disagree with it. The full per-dataset results, including MATH-500, StrategyQA and batch-1 latency, are in test_results.json; training_config.json, provenance.json and export_verification.json record the training settings, source checkpoint hashes and export checks.

Related models

• CIRF-Qwen3-8B-Instruct-Full (this model)
• CIRF-Qwen3-8B-Instruct-Zero-Results
• CIRF-Qwen3-8B-Instruct-Backward-Greedy-3
• CIRF-Qwen3-8B-Instruct-Backward-Greedy-6
• CIRF-Qwen3-8B-Instruct-Backward-Greedy-12
• CIRF-Qwen3-8B-Instruct-Backward-Greedy-18
• CIRF-Qwen3-8B-Instruct-Backward-Greedy-24

License

The base model is Qwen/Qwen3-8B, released under Apache 2.0; this derivative keeps that license (LICENSE). Changes comprise the CIRF functional-token vocabulary, learned embedding and output rows, supervised adaptation, and the merged LoRA update.